# cbloom rants

## 12/11/2010

### 12-11-10 - Perceptual Notes of the Day

You have to constantly be aware of how you're pooling data. There's no a-priori reason to prefer one way or the other; when a human looks at an image do they see the single worst spot, or do they see the average? If you show someone two images, one with a single very bad error, and another with many small errors, which one do they say is worse?

There are also various places to remap scales. Say you have measured a quantity Q in two images, one way to make an error metric out of that is :

```
err = Lp-norm {  remap1(  remap2( Q1 ) - remap2( Q2 ) )  }

```
Note that there are two different remapping spots - remap1 converts the delta into perceptually linear space, while remap2 converts Q into space where it can be linearly delta'd. And then of course you have to tweak the "p" in Lp-norm.

You're faced with more problems as soon as you have more quantities that you wish to combine together to make your metric. Combine them arithmetically or geometrically? Combine before or after spatial pooling? Weight each one in the combination? Combine before or after remapping? etc.

This is something general that I think about all the time, but it also leads into these notes :

For pixel values you have to think about how to map them (remap2). Do you want light linear? (gamma decorrected) or perhaps "perceptually uniform" ala CIE L ?

The funny thing is that perception is a mapping curve that's opposing to gamma correction. That is, the monitor does something like signal ^ 2.2 to make light. But then the eye does something like light ^ (1/3) (or perhaps more like log(light)) to make mental perception. So the net mapping of pixel -> perceptuion is actually pretty close to unity.

On the subject of Lp norms, I wrote this in email (background : a while ago I found that L1 error was much better for my video coder than L2) :

```
I think L1 was a win mainly because it resulted in better R/D bit distribution.

In my perceptual metric right now I'm using L2 (square error) + masking , which appears to be better still.

I think that basically L1 was good because if you use L2 without masking, then you care about errors in noisy parts of the image too much.

That is, say your image is made of two blocks, one is very smooth, one has very high variation (edges and speckle).

If you use equal quantizers, the noisy block will have a much larger error.

With L2 norm, that means R/D will try very hard to give more bits to the noisy block, because changing 10 -> 9 helps more than changing 2 -> 1.

With L1 norm it balances out the bit distribution a bit more.

But L2-masked norm does the same thing and appears to be better than L1 norm.

```

Lastly a new metric :

PSNR-HVS-T is just "psnr-hvs" but with thresholds. (*3) In particular, if PSNR-HVS is this very simple pseudo-code :

```
on all 8x8 blocks
dct1,dct2 = dct of the blocks, scaled by 1/ jpeg quantizers

for i from 0 -> 63 :
err += square( dct1[i] - dct2[i] )

```
then PSNR-HVS-T is :
```
on all 8x8 blocks
dct1,dct2 = dct of the blocks, scaled by 1/ jpeg quantizers

err += w0 * square( dct1 - dct2 )

for i from 1 -> 63 :
err += square( threshold( dct1[i] - dct2[i] , T ) )

```
the big difference between this and "PSNR-HVS-M" is that T is just a constant, not the complex adaptive mask that they compute (*1).

```
threshold is :

threshold(x,t) = MAX( (x - T), 0 )

```

The surprising thing is that PSNR-HVS-T is almost as good as PSNR-HVS-M , which is in turn basically the best perceptual metric (*2) that I've found yet. Results :

```
PSNR-HVS-T Spearman : 0.935155

fit rmse :
PSNR-HVS   : 0.583396  [1,2]
VIF        : 0.576115  
PSNR-HVS-M : 0.564279  
IW-MS-SSIM : 0.563879  
PSNR-HVS-T : 0.557841
PSNR-HVS-M : 0.552151  

```
In particular, it beats the very complex and highly tweaked out IW-MS-SSIM. While there are other reasons to think that IW-MS-SSIM is a strong metric (in particular, the maximum difference competition work), the claims that they made in the papers about superiority of fitting to MOS data appear to be bogus. A much simpler metric can fit better.

footnotes :

*1 = note on the masking in PSNR-HVS-M : The masked threshold they compute is proportional to the L2 norm of all the DCT AC coefficients. That's very simple and unsurprising, it's a simple form of variance masking which is well known. The funny thing they do is multiply this threshold by a scale :

```
scale = ( V(0-4,0-4) + V(0-4,4-8) + V(4-8,0-4) + V(4-8,4-8) ) / ( 4*V(0-8,0-8) )

V(x,y) = Variance of pixels in range [x,y] in the block

threshold *= sqrt( scale );

```
that is, divide the block into four quadrants; take the variance within each quadrant and divide by the variance of the whole block. What this "scale" does is reduce the threshold in areas where the AC activity is caused by large edges. Our basic threshold was set by the AC activity, but we don't want the important edge shapes to get crushed, so we're undoing that threshold when it corresponds to major edges.

Consider various extremes; if the block has no large-scale structure, only random variation all over, then the variance within each quad will equal the variance of the whole, which will make scale ~= 1. On the other hand, if there's a large-scale edge, then the variance on either side of the edge may be quite small, so at least one of the quads has a small variance, and the result is that the threshold is reduced.

Two more subtle follows up to this :

1A : Now, the L2 sum of DCT AC coefficients should actually be proportional to the variance of the block ! That means instead of doing

```
threshold^2 ~= (L2 DCT AC) * (Variance of Quads) / (Variance of Blocks)

```
we could just use (Variance of Quads) directly for the threshold !! That is true, however when I was saying "DCT" earlier I meant "DCT scaled by one over JPEG quantizer" - that is, the DCT has CSF coefficients built into it, which makes it better than just using pixel variances.

1B : Another thought you might have is that we are basically using the DCT L2 AC but then reducing it for big edges. But the DCT itself is actually a decent edge detector, the values are right there in DCT[0,1] and DCT[1,0] !! So the easiest thing would be to just not include those terms in the sum, and then don't do the scale adustment at all!

That is appealing, but I haven't found a formulation of that which performs as well as the original complex form.

(*2) = "best" defined as able to match MOS data, which is a pretty questionable measure of best, however it is the same one that other authors use.

(*3) = as usual, when I say "PSNR-HVS" I actually mean "MSE-HVS", and actually now I'm using RMSE because it fits a little better. Very clear, I know.

## 12/09/2010

### 12-09-10 - Rank Lookup Error

Trying some other methods of testing to make sure the function fit isn't screwing me up too much.

Spearman rank correlations (it's just the Pearson correlation on sort ranks) :

```
mse.txt                 0.66782
square_ms_ssim_y.txt    0.747733
scielabL1.txt           0.855355
scielabL2.txt           0.868207
vif.txt                 0.87631
ms_ssim.txt             0.89391
iw_ms_ssim.txt          0.901085
wsnr.txt                0.90905
mydctdelta.txt          0.932998
my_psnr_hvs_m.txt       0.94082
mydctdeltanew.txt       0.944086

```

I just thought of a new way to check the fit for this kind of scenario which I think is pretty cool.

You have two vectors of values. One vector is the original which you wish to match, the other is output by your program to try to match the original vector. The problem is that your program outputs in some other scale, different units which may involve some warping of the curve, and you don't know what it is. You wish to find how close your program is to the target without worry about matching that curve.

Well, one way is "rank lookup error" :

```
given vectors "orig" and "test"

find "test_rank" such that
r = test_rank[i] means item i is the r'th in sorted order in the vector test

find "sorted_orig" = orig, sorted

Sum{i} :
err += square( orig[i] - sort_orig[ test_rank[ i ] ] )

```
that is, the fit value for mapping from test's scale to orig is to find the sort index within test, and lookup the value in the sorted list of originals.

Obviously this isn't quite ideal; it does handle ties and near-ties pretty well though (better than Spearman/Kendall, because you get less error contribution when you get the rank wrong of two items with very similar value). Most importantly it avoids all the fidgetty function fitting stuff.

Here are the scores with "rank lookup rmse" :

```
mse.txt                 1.310499
square_ms_ssim_y.txt    1.090392
scielabL1.txt           0.853738
scielabL2.txt           0.835508
ms_ssim.txt             0.667152
wsnr.txt                0.643821
vif.txt                 0.591508
iw_ms_ssim.txt          0.575474
mydctdelta.txt          0.548946
my_psnr_hvs_m.txt       0.543057
mydctdeltanew.txt       0.494918

```
if nothing else it's a good sanity check for the fitting stuff.

Also with rank-lookup-error you don't have to worry about transforms like acos for ssim or undo-psnr or anything else that's monotonic.

For comparison, these were the fit scores :

```
mse.txt                 1.169794
square_ms_ssim_y.txt    1.005849
scielabL1.txt           0.819501
scielabL2.txt           0.808595
ms_ssim.txt             0.639193
wsnr.txt                0.632670
vif.txt                 0.576114
iw_ms_ssim.txt          0.563879
my_psnr_hvs_m.txt       0.552151
mydctdelta.txt          0.548938
mydctdeltanew.txt       0.489756

```

mydctdeltanew is a new one; it just uses mydctdelta for only the AC part of the DCT (excludes the DC which is gross luma differences), and then it adds back on the gross luma difference using a form similar to IW-MS-SSIM (that is, using large filters and then using both variance masking and saliency boosting).

### 12-09-10 - Perceptual vs TID

My TID2008 score is the RMSE of fitting to match MOS values (0-9) ; fit from perceptual metric to MOS score is of the form Ax + Bx^C , so 0 -> 0 , and for fitting purposes I invert and normalize the scale so that 0 = no error and 1 = maximum error. For fitting I try {value},{acos(value)}, and {log(value)} and use whichever is best.

I also exclude the "exotic" distortions from TID because they are weird and not something well handled by most of the metrics (they favor SSIM which seems to be one of the few that handles them okay). I believe that including them is a mistake because they are very much unlike any distortion you would ever get from a lossy compressor.

I also weight the MOS errors by 1/Variance of each MOS ; basically if the humans couldn't agree on a MOS for a certain image, there's no reason to expect the synthetic metric to agree. (this is also like using an entropy measure for modeling the human data; see here for example)

```
MSE        : 1.169795  [1,2]
MSE-SCIELAB: 0.808595  
MS-SSIM    : 0.691881  [1,2]
SSIM       : 0.680292  
MS-SSIM-fix: 0.639193   (*)
WSNR       : 0.635510  
PSNR-HVS   : 0.583396  [1,2] (**)
VIF        : 0.576115  
PSNR-HVS-M : 0.564279   (**)
IW-MS-SSIM : 0.563879  
PSNR-HVS-M : 0.552151   (**)
MyDctDelta : 0.548938  

 = scores from TID reference "metrics_values"
 = scores from me
[1,2] = confirmed scores from me and TID are the same

```
the majority of these metrics are Y only, using rec601 Y (no gamma correction) (like metric mux).

WARNING : the fitting is a nasty evil business, so these values should be considered plus/minus a few percent confidence. I use online gradient descent (aka single layer neural net) to find the fit, which is notoriously tweaky and sensitive to annoying shit like the learning rate and the order of values and so on.

(*) = MS-SSIM is the reference implementation and I confirmed that I match it exactly and got the same score. As I noted previously , the reference implementation actually uses a point subsample to make the multiscale pyramid. That's obviously a bit goofy, so I tried box subsample instead, and the result is "MS-SSIM-fix" - much better.

(**) = PSNR-HVS have received a PSNR-to-MSE conversion (ye gods I hate PSNR) ; so my "PSNR-HVS-M" is actually "MSE-HVS-M" , I'm just sticking to their name for consistency. Beyond that, for some reason my implementation of PSNR-HVS-M () is better than theirs () and I haven't bothered to track down why exactly (0.564 vs 0.552).

WSNR does quite well and is very simple. It makes the error image [orig - distorted] and then does a full image FFT, then multiplies each tap by the CSF (contrast sensitivity function) for that frequency, and returns the L2 norm.

PSNR-HVS is similar to WSNR but uses 8x8 DCT's instead of full-image FFT. And of course the CSF for an 8x8 DCT is just the classic JPEG quantization matrix. That means this is equivalent to doing the DCT and scaling by one over the JPEG quantizers, and then taking the L2 delta (MSE). Note that you don't actually quantize to ints, you stay in float, and the DCT should be done at every pixel location, not just 8-aligned ones.

VIF is the best of the metrics that comes with TID ; it's extremely complex, I haven't bothered to implement it or study it that much.

PSNR-HVS-M is just like PSNR-HVS, but adds masked thresholds. That is, for each tap of the 8x8 DCT, a just noticeable threshold is computed, and errors are only computed above the threshold. The threshold is just proportional to the L2 AC sum of the DCT - this is just variance masking. They do some fiddly shit to compensate for gross variance vs. fine variance. Notably the masking is only used for the threshold, not for scaling above threshold (though the CSF still applies to scaling above threshold).

IW-MS-SSIM is "Information Weighted" MS-SSIM. It's a simple weight term for the spatial pooling which makes high variance areas get counted more (eg. edges matter). As I've noted before, SSIM actually has very heavy variance masking put in (it's like MSE divided by Variance), which causes it to very severely discount errors in areas of high variance. While that might actually be pretty accurate in terms of "visibility", as I noted previously - saliency fights masking effects - that is, the edge areas are more important to human perception of quality. So SSIM sort of over-crushes variance and IW-SSIM puts some weight back on them. The results are quite good, but not as good as the simple DCT-based metrics.

MyDctDelta is some of the ideas I wrote about here ; it uses per-pixel DCT in "JPEG space" (that is, scaled by 1/JPEG Q's), and does some simple contrast-band masking, as well as multi-scale sum deltas.

There's a lot of stuff in MyDctDelta that could be tweaked that I haven't , I have no doubt the score on TID could easily be made much better, but I'm a bit afeared of overtraining. The TID database is not big enough or varied enough for me to be sure that I'm stressing the metric enough.

Another aside about SSIM. The usual presentation says compute the two terms :

```
V = (2*sigma1*sigma2 + C2)/(sigma1_sq + sigma2_sq + C2);
C = (sigma12 + C3)/(sigma1*sigma2 + C3);

```
for variance and correlation dot products. We're going to wind up combining these multiplicatively. But then notice that (with the right choice of C3=C2/2)
```
V*C = (2*sigma12 + C2/2)/(sigma1_sq + sigma2_sq + C2);

```
so we can avoid computing some terms.

But this is a trick. They've actually changed the computation, because they've changed the pooling. The original SSIM has three terms : mean, variance, and correlation dot products. They are each computed on a local window, so you have the issue of how you combine them into a single score. Do you pool each one spatially and then cross-multiply? Or do you cross-multiply and then pool spatially ?

```
SSIM = Mean{M} * Mean{V} * Mean{C}

or

SSIM = Mean{M*V*C}

```
which give quite different values. The "efficient" V*C SSIM winds up using :
```
SSIM = Mean{M} * Mean{V*C}

```
which is slightly worse than doing separate means and multiplying at the end (which is what MS-SSIM does).

## 12/07/2010

### 12-07-10 - Patents

Ignacio wrote about software patents a while ago and it's got me thinking.

First of all I applaud the general idea that every individual is responsible for acting morally, regardless of the environment they live in. I see far too many people who use the broken system as an excuse to behave badly themselves. eg. lots of people are polluting it doesn't matter how much I hurt the environment, the tax system is broken it doesn't matter how much I cheat, the patent system sucks so I have to be a part of it, etc. I believe this attitude is a facile rationalization for selfish behavior.

In any case, let's get back to the specific topic of patents.

In my youth I used to be the lone pro-patent voice in a sea of anti-patent peers. Obviously I was against the reality of the patent system, which consists of absurd over-general patents on things that aren't actually innovations, and the fact that expert review of technical issues before a court is an absurd farce in which money usually wins. But I was pro the basic idea of patents, partly for purely selfish reasons, because I had these technical inventions I had made and was hoping I could somehow get rich from them, as I saw others do. As a young individual inventor, I believed that getting a patent was my only way to get a fair price for my work.

Now I have a different view for a few reasons :

1. Policy should be made based on the reality of an issue, not your theoretical ideal.

The reality is that patents (and particularly software patents) are ridiculously broken. The court system does not have the means to tell when patents are reasonable or not, and it is unrealistic to think that that can be fixed.

2. The purpose of laws should be to ensure the greatest good for society.

Even if you think software patents are good for the lonely independent developer, that in itself is not reason enough to have them. You have to consider the net benefit to society. I believe that the world would be better off without software patents, but this is a little tricky.

What are the pro-patent arguments ?

One is that they encourage research funding, that companies wouldn't spend money on major research if they didn't have the patent system to ensure a monopoly for their invention.

I find this argument generally absurd. Do you think that IBM or Microsoft really wouldn't fund research that they believe will improve their business if they couldn't patent the result? Companies will fund research any time it is likely profitable; a long-term monopoly from a patent doesn't really change the research equation from "not profitable" to "profitable", it changes it to "ridiculously profitable".

Furthermore, in reality, most of the major tech companies don't actually use their patents to keep monopolies on technology, rather they engage in massive cross-license agreements to get open access to technologies. Patents wind up being a huge friction and cost for these companies, and you have to maintain a big war chest of your own patents to ensure that you can participate in the cross-licensing. The end result of this is yet another oligarchy. The big tech companies form cross-license agreements, and smaller players are frozen out. This is a huge friction to free market innovation.

Now, I believe one legitimate point is that in a world without patents, companies might be more motivated to keep their innovations secret. One pro-patent argument is that it allows companies to patent their work and then publish without fear of losing it. Of course, that is a bit of a false promise, because the value to the public of getting a publication which describes a patented algorithm is dubious. (yes obviously there is some value because you get to read about it, but you don't actually get to use it!).

Finally it is absolutely offensive to me that researchers who receive public funding are patenting their works, or that university professors patent their works. If you receive any public funding, your work should be in the public domain (and it should not be published in the pay-for-access journals of the ACM or IEEE).

But this is an issue that goes beyond software. Are any patents good for society? Certainly medical patents have encouraged lots of expensive research in recent years, and this is often used as a pro-patent argument. But lots of those new expensive drugs have been shown to be no better than their cheap predecessors. Certainly patents provide a massive incentive for drug companies to push the new expensive monopolized product on you, which is a bad effect for society. Would you actually have significantly less useful research if there were no patents? Well, 30-40% of medical research is publicly funded, so that wouldn't go away, and without patents that publicly funded research would be much more efficient, because they could be open and not worry about infringement, and furthermore it would be more focused on research that provides tangible benefits as opposed to research that leads to profits. It's a complicated issue, but it's definitely not obvious that the existance patents actually improve the net quality of medical treatment.

In summary, I believe that patents do accomplish some good, but you have to weigh that against the gains you would get if you had no patents. I believe the good from no patents is greater than the good from patents.

In any case, hoping for patents to go away is probably a pipe dream.

Smaller goals are these :

1. I find it absolutely sick that public universities are patenting things. That needs to stop. Professors/researchers need to take the lead by refusing to patent their inventions.

Any corporation that receives the bullshit "R&D" tax break should be required to make all their patents public domain. Anyone that gets a DoD or NSF research grant should be required to make the results of their work public domain. How can you justify taking public money for R&D and then locking out the public from using the results?

2. Some rich charity dude should create the "public patent foundation" whose goal is to supports the freedom of ideas, and has the big money to fight bullshit patents in court. The PPF could also actively work to publish prior art and even in some cases to apply for patents, which would then be officially released into the public domain.

A more extreme idea would be to make the PPF "viral" like the GPL - build up a big war chest of patents, and then release them all for free use - but only to other people who release their own patents under the same license. All the PPF has to do is get a few important patents and it can force the opening of them all.

(deja vu , I just realized I wrote this before )

## 12/06/2010

### 12-06-10 - More Perceptual Notes

To fit an objective synthetic measure to surveyed human MOS scores, the VQEG uses the form :
```
A1 + A2 * x + A3 * logistic( A4 * x + A5 )

```
I find much better fits from the simpler form :
```
B1 + B2 * x ^ B3

```

Also, as I've mentioned previously, I think the acos of SSIM is a much more linear measure of quality (SSIM is like a Pearson correlation, which is like a dot product, so acos makes it like an angle). This is borne out by fitting to TID2008 - the acos of SSIM fits better to the human scores (though fancy fit functions can hide this - this is most obvious in a linear fit).

But what's more, the VQEG form does not preserve the value of "no distortion", which I think is important. One way to do that would be to inject a bunch of "zero distortion" data points into your training set, but a better way is to use a fit form that ensures it explicitly. In particular, if you remap the measured MOS and your objective score such that 0 = perfect and larger = more distorted, then you can use a fit form like :

```
C1 * x + C2 * x ^ C3

```
(C3 >= 0) , so that 0 maps to 0 absolutely, and you still get very good fits (in fact, better than the "B" form with arbitrary intercept). (note that polynomial fit (ax+bx^2+cx^3) is very bad, this is way better).

A lot of people are using Kendall or Spearman rank correlation coefficients to measure perceptual metrics (per VQEG recommendation), but I think that's a mistake. The reason is that ranking is not what we really care about. What we want is a synthetic measure which correctly identifies "this is a big difference" vs. "this is a small difference". If it gets the rank of similar differences a bit wrong, that doesn't really matter, but it does show up as a big difference in Kendall/Spearman. In particular, the *value* difference of scores is what we care about. eg. if it should have been a 6 and you guess 6.1 that's no big deal, but if you guess it's a 9 that's very bad. eg. if your set of MOS scores to match is { 3, 5.9, 6, 6.1, 9 } , then getting the middle 3 in the wrong order is near irrelevant, but the Kendall & Spearman have no accounting for that, they are just about rank order.

The advantage of rank scores of course is that you don't have to do the functional fitting stuff above, which does add an extra bias, because some objective scores might work very well with that particular functional fit, while others might not.

## 12/02/2010

### 12-02-10 - Perceptual Metric Rambles of the Day

There's an interesting way in which saliency (visual attention) fights masking (reduction of sensitivity due to local variance). People who weight image metrics by saliency give *more* importance to high contrast areas like edges. People who weight image metrics by masking give *less* importance to high contrast areas. Naively it seems like those would just cancel out.

It's subtle. For one thing, a smooth area is only not "salient" if it's smooth in both the original and the disorted image. If you do something like creating block artifacts, then those block edges are salient (high visual attention factor). Also, they work at sort of different scales. Saliency works on intermediate frequencies - the brain specifically cares about detail at a certain scale that detects edges and shapes, but not speckle/texture and not gross offsets. Contrast or variance masking applies mainly to the aditional frequencies in an area that has a strong medium-scale signal.

MS-SSIM actually has a lot of tweaked out perceptual hackiness. There's a lot of vision modelling implicit in it. They use a Gaussian window with a tweaked out size - this is a model of the spatial filter of the eye. The different scales have different importance contributions - this is a model of the CSF, variable sensitivity and different scales. The whole method of doing a Pearson-style correlation means that they are doing variance masking.

The issue of feeding back a perceptual delta into a non-perceptual compressor is an interesting one.

Say you have a compressor which does R-D optimization, but its D is just MSE. You can pretty easily extend that to do weighted R-D, with a weight either per-pixel or per-block. Initially all the weights are the same (1.0). How do you adjust the weights to optimize an external perceptual metric?

Assume your perceptual metric returns not just a number but an error map, either per pixel or per block.

First of all, what you should *not* do is to simply find the parts with large perceptual error and increase their weight. That would indeed cause the perceptual error at those spots to reduce - but we don't know if that is actually a good thing in an R-D sense. What that would do is even out the perceptual error, and evening it out is not necessarily the best R-D result. This is, however, an easy to get the minimax perceptual error!

```
Algorithm Minimax Perceptual Error :

1. Initial W_i = 1.0
2. Run weighted R-D non-perceptual compressor (with D = weighted MSE).
3. Find perceptual error map P_i
4. W_i += lambda * P_i
(or maybe W_i *= e^(lambda * P_i))
5. goto 2

```
what about minimizing the total perceptual error?

What you need to do is transform the R/D(MSE) that the compressor uses into R/D(percep). To do that, you need to make W_i ~= D(percep)/D(mse).

If the MSE per block or pixel is M_i, the first idea is just to set W_i = ( P_i / M_i ) . This works well if P and M have a near-linear relationship *locally* (note they don't need to be near-linear *globally* , and the local linear relationship can vary from one location to another).

However, the issue is that the slope of P_i(R) may not be near linear to M_i(R) (as functions of rate) over its entire scale. Obviously for small enough steps, they are always linear.

So, you could run your compress *twice* to create two data points, so you have M1,M2 and P1,P2. Then what you want is :

```
P ~= P1 + ((P2 - P1)/(M2 - M1)) * (M - M1)

slope := (P2 - P1)/(M2 - M1)

P ~= slope * M + (P1 - slope * M1)

C := (P1 - slope * M1)

P ~= slope * M + C

I want J = R + lambda * P

use

J = R + lambda * (slope * M + C)
J = R + lambda * slope * M + lambda * C

so

W_i = slope_i

J = R + lambda * W * M

```
with the extra lambda * C term ; C is just a constant per block, so it can be ignored for R-D optimization on that block. If you care about the overall J of the image, then you need the sum of all C's for that, which is just :
```
C_tot = Sum_i (P1_i - slope_i * M1_i)

```
it's obviously just a bias to turn our fudged MSE-multiplied-by-slope into the proper scale of the perceptual error.

I think this is pretty strong, however it's obviously iterative, and you must keep the steps of the iteration small - this is only valid if you don't make very big changes in each step of the iteration.

There's also a practical difficulty that it can be hard to actually generate different M1 and M2 on some blocks. You want to generate both M1 and M2 very near the settings that you will be using for the final compress, because large changes could take you to vastly different perceptual areas. But, when you make small changes to R, some parts of the image may change, but other parts might not actually change at all. This gives you M1=M2 in some spots which is a divide by zero, which is annoying.

In any case I think something like this is an interesting generic way to adapt any old compressor to be perceptual.

Note that this kind of "perceptual importance map" generation cannot help your coder make internal decisions based on perceptual optimization - it can only affect bit allocation. That is, it can help a coder with variable bit allocation (either through truncation or variable quantization) to optimize the bit allocation.

In particular, a coder like H264 has lots of flexibility; you can choose a macroblock mode, an inter block can choose a movec, an intra block can choose a prediction mode, you can do variable quantizers, and you can truncate coefficients. This kind of iterative approach will not help the mode decisions to pick modes which have less perceptual error, but it will help to change the bit allocation to give more bits to blocks where it will help the preceptual error the most.

### 12-02-10 - Orphaned Tech Ideas

If you have a really good idea, you have to just go for it right away. Sometimes I have an idea, but I'm busy with something else, so I just write it down, and then try to come back to it later, but when you come back the excitement is gone. You have to pounce on that moment of inspiration.

That said, there are many things I'd like to do right now but probably have to just set aside and never come back to :

DXTC for lightmaps (and other non-photographic data). In particular there should be a way to preserve smooth data with much higher precision.

DXTC optimized for perceptual metrics. Should be easy, I'll probably do this.

cblib::hash_table entry currently always has {hash, key, data} , should change that to a functor with calls { hash(), key(), data() } so that user can choose to store them or not. (in many cases you don't need seperate key and data because the key is the data).

csv util suite. csv -> google chart, csv lsqr fit, csv eval, etc. basically matlab in command line utils that work on csv's. Would let me do processing from batch files and make html logs of compression runs and such.

Texture synthesis from examples. Fast enough for realtime so it can be used with SVT. Should be progressive. I believe this is the right way to do large world texturing. Texture synthesis from procedural/perlin noise stuff is too hard for artists to use, and doing fancy synthesis offline and then storing it ala Id just strikes me as profoundly silly.

Really good JPEG decoder. Maximum-likelihood decompression. Nosratinia and POCS. You can basically remove all blocking/ringing artifacts. This seems to me like a very useful and important piece of software and I'm surprised it doesn't exist.

Luma aided chroma upsample. Could be part of "Really good JPEG decoder" ; also useful for improving video decompression quality.

Finish lossy image comparison test and release the exe which autogens nice HTML reports.

## 11/30/2010

### 11-30-10 - Tchebychev

A little utility, and example of how I use templates : (this is not meant to be efficient, it's for offline apps, not realtime use)
```
template < int n >
struct Tchebychev
{
double operator () ( double x )
{
return 2.0 * x * Tchebychev < n-1>()(x) - Tchebychev < n-2>()(x);
}
};

template < >
struct Tchebychev < 0 >
{
double operator () ( double x )
{
return 1.0;
}
};

template < >
struct Tchebychev < 1 >
{
double operator () ( double x )
{
return x;
}
};

double TchebychevN(int n,double x)
{
switch(n) {
case 0 : return Tchebychev<0>()(x);
case 1 : return Tchebychev<1>()(x);
case 2 : return Tchebychev<2>()(x);
case 3 : return Tchebychev<3>()(x);
case 4 : return Tchebychev<4>()(x);
case 5 : return Tchebychev<5>()(x);
case 6 : return Tchebychev<6>()(x);
case 7 : return Tchebychev<7>()(x);
case 8 : return Tchebychev<8>()(x);
case 9 : return Tchebychev<9>()(x);
NO_DEFAULT_CASE
}
return 0;
}

template < typename functor >
void FindTchebychevCoefficients(functor f,double lo = 0.0, double hi = 1.0, int N = (1<<20))
{
double PI_over_N = PI/N;

#define NUM_T   6

double t[NUM_T] = { 0 };

for(int k=0;k < N;k++)
{
double pk = PI_over_N * (k + 0.5);
double x_k = cos(pk);
double farg = lo + (hi - lo) * 0.5 * (x_k+1.0);
double fval = f( farg );
for(int j=0;j < NUM_T;j++)
{
t[j] += fval * cos( pk * j );
}
}
for(int j=0;j < NUM_T;j++)
{
t[j] *= 1.0 / N;
if ( j != 0 )
t[j] *= 2.0;

//lprintfvar(t[j]);
lprintf("t %d: %16.9f , %16.8g\n",j,t[j],t[j]);
}

double errSqr[NUM_T] = { 0 };
double errMax[NUM_T] = { 0 };
for(int i=0;i < N;i++)
{
double xunit = (i+0.5)/N;
double farg = lo + (hi - lo) * xunit;
double xt = xunit*2.0 - 1.0;
double y = f(farg);
double p = 0.0;
for(int j=0;j < NUM_T;j++)
{
p += t[j] * TchebychevN(j,xt);
errSqr[j] += fsquare(y-p);
errMax[j] = MAX( errMax[j] , fabs(y-p) );
}
}
for(int j=0;j < NUM_T;j++)
{
lprintf("%d : errSqr = %g errMax = %g\n",j,errSqr[j]/N,errMax[j]);
}
}

```

You can also very easily find inner products by using the Integrate<> template I posted earlier plus this simple product adaptor :

```
template < typename f1, typename f2 >
struct FunctorProduct
{
double operator () ( double x )
{
return f1()(x) * f2()(x);
}
};

```

(eg. you can trivially find Hermite or Legendre series this way by doing inner products).

Another handy helper is a range remapper :

```
template < typename functor >
struct RemapFmTo
{
double m_fmLo; double m_fmHi;
double m_toLo; double m_toHi;

RemapFmTo(
double fmLo, double fmHi,
double toLo, double toHi )
: m_fmLo(fmLo), m_fmHi(fmHi), m_toLo(toLo), m_toHi(toHi)
{
}

double operator () ( double x )
{
double t = (x - m_fmLo) / (m_fmHi - m_fmLo);
double y = t * (m_toHi - m_toLo) + m_toLo;
return functor()(y);
}
};

template < typename functor >
struct RemapUnitTo
{
double m_toLo; double m_toHi;

RemapUnitTo(
double toLo, double toHi )
:  m_toLo(toLo), m_toHi(toHi)
{
}

double operator () ( double x )
{
double y = x * (m_toHi - m_toLo) + m_toLo;
return functor()(y);
}
};

```

Now we can trivially do what we did before to find the optimal approximation in a known range :

```
struct SqrtOnePlusX
{
double operator () ( double x )
{
return sqrt( 1 + x );
}
};

RemapUnitTo< SqrtOnePlusX > tf(-0.075f,0.075f);
FindTchebychevCoefficients( tf );

```
and the output is :
```
t 0:      0.999647973 ,       0.99964797
t 1:      0.037519818 ,      0.037519818
t 2:     -0.000352182 ,   -0.00035218223
t 3:      0.000006612 ,   6.6121459e-006
t 4:     -0.000000155 ,  -1.5518238e-007
t 5:      0.000000004 ,   4.0790168e-009
0 : errSqr = 0.000469204 errMax = 0.0378787
1 : errSqr = 5.78832e-008 errMax = 0.000358952
2 : errSqr = 2.12381e-011 errMax = 6.77147e-006
3 : errSqr = 1.18518e-014 errMax = 1.59378e-007
4 : errSqr = 8.23726e-018 errMax = 4.19794e-009
5 : errSqr = 6.54835e-021 errMax = 1.19024e-010

```
I guess Remez minimax polynomials are better, but this is so easy and it gives you a good starting point, then you can just numerically optimize after this anyway.

obviously the TchebychevN dispatcher sucks and is silly, but you don't actually use it in practice; you know which polynomials you want and you use something like :

```
double approx =
c0 * Tchebychev<0>(x) +
c1 * Tchebychev<1>(x) +
c2 * Tchebychev<2>(x) +
c3 * Tchebychev<3>(x);

```
which the compiler handles quite well.

I also did Legendre polynomials :

```
template < int n >
struct Legendre
{
double operator () ( double x )
{
double L_n1 = Legendre < n-1 >()(x);
double L_n2 = Legendre < n-2 >()(x);
const double B = (n-1)/(double)n; // n-1/n
return (1+B)*x*L_n1 - B*L_n2;
}
};

template< > struct Legendre<0>
{
double operator () ( double x ) { return 1.0; }
};

template< > struct Legendre<1>
{
double operator () ( double x ) { return x; }
};

// but ryg's loop variants are mostly just better :

__forceinline double legendre_ryg(int n, double x)
{
if (n == 0) return 1.0;
if (n == 1) return x;

double t = x, t1 = 1, t2;
for (int i=1; i < n; i++) {
t2 = t1;
t1 = t;
const double B = (i)/(double)(i+1);
t = (1+B)*x*t1 - B*t2;
}
return t;
}

__forceinline double chebyshev_ryg(int n, double x)
{
if (n == 0) return 1.0;
if (n == 1) return x;

double t = x, t1 = 1, t2;
for (int i=1; i < n; i++) {
t2 = t1;
t1 = t;
t = 2.0 * x * t1 - t2;
}
return t;
}

// you can find Legendre coefficients like so :

RemapUnitTo< Legendre<0> > ShiftedLegendre0(-1,1);

double c0 = Integrate(0.0,1.0,steps,MakeFunctorProduct(functor(),ShiftedLegendre0));
// (don't forget normalization)

// using new helper :

template < typename f1, typename f2 >
FunctorProduct< f1,f2 > MakeFunctorProduct( f1 _f1 , f2 _f2 ) { return FunctorProduct< f1,f2 >(_f1,_f2); }

```

### 11-30-10 - Reference

You can do weighted least squares using a normal least squares solver. To solve :
```
A x = b

```
in the least squares sence (that is, minimize |Ax - b|^2) , with weights W , you just are minimizing the error :
```
E = Sum_i { W_i * ( Sum_j {A_ij * x_j} - b_i )^2 }

E = Sum_i { ( Sum_j { sqrt(W_i) * A_ij * x_j} - sqrt(W_i) * b_i )^2 }

so set

A'_ij = A_ij * sqrt(W_i)
b'_i = b_i * sqrt(W_i)

then

E = Sum_i { ( Sum_j { A'_ij * x_j } - b'_i )^2 }

```
so we can just do a normal lsqr solve for x using A' and b'.

If you have a Gaussian random event, probability of x is P(x), the (natural) bits to code an event is -ln(P(x)) =

```
P(x) = 1/(sigma * sqrt(2pi)) * e^( - 0.5 * ((x-c)/sigma)^2 )

H = -ln(P(x)) = ln(sigma * sqrt(2pi)) + 0.5 * ((x-c)/sigma)^2

H = const + ln(sigma) + 0.5 * ((x-c)/sigma)^2

```
And bits to code events is a good measure of how well events fit a model, which is useful for training a model to observed events. In particular if we ignore the constants and the ln term,
```
total H = Sum_ij { (1/sigma_i^1) * (x_j - c_i)^2 }

```
the bit cost measure is just L2 norm weighted by one over sigma^2 - which is intuitive, matching the prediction of very narrow Gaussians is much more important than matching the prediction of very wide ones.

Now that I have autoprintf, I find this very handy :

```
#define lprintfvar(var) lprintf(STRINGIZE(var) " : ",var,"\n");

```

## 11/29/2010

### 11-29-10 - Useless hash test

Clocks to do the full hash computation and lookup on char * strings of english words : ```
hash functions :

0 : FNV1A_Hash_Jesteress
1 : FNV1A_HashStr_WHIZ
2 : FNVHashStr
3 : HashFNV1
4 : MurmurHash2_Str
5 : HashKernighanRitchie
6 : HashBernstein
7 : HashLarson
8 : Hash17
9 : Hash65599
10 : stb_hash
11 : HashWeinberger
12 : HashRamakrishna
13 : HashOneAtTime
14 : HashCRC

```

The hash functions are not given the string length, they must find it. For hashes that work on dwords the fast way to do that is using

```
#define DWORD_HAS_ZERO_BYTE(V)       (((V) - 0x01010101UL) & ~(V) & 0x80808080UL)

```

Hash functions are all inlined of course. The easy way to do that is to use macros to make functors for your hash functions and then pass them into a templated hash test function, something like this :

```
#define TEST_HT(hashfunc) \
struct STRING_JOIN(htops_,hashfunc) : public hash_table_ops_charptr \
{ inline hash_type hash_key(const char_ptr & t) { return hashfunc( t ); } }; \
typedef hash_table < char_ptr,Null,STRING_JOIN(htops_,hashfunc) > STRING_JOIN(ht_,hashfunc); \
test_a_hash < STRING_JOIN(ht_,hashfunc)>(STRINGIZE(hashfunc),words,queries);

TEST_HT(FNV1A_HashStr_WHIZ);
TEST_HT(FNVHashStr);

```

This is using cblib:hash_table which is a reprobing hash table, with the hash values in the table, quadratic reprobing, and with special key values for empty & deleted (not a separate flag array). cblib:hash_table is significantly faster than khash or std::hash_map (though it takes more memory than khash and other similar hash tables due to storing the hash value). (BTW this was studied before extensively; see earlier blog posts; the four major deviations from khash are all small wins : 1. don't use separate flags because they cause an extra cache miss, 2. use cache-friendly quadratic reprobe, not rehashing, 3. use pow2 table size not prime, 4. store hash value).

The four tests in the chart above are :

```
blue:   StringHash_Test("M:\\CCC\\canterbury\\LCET10.txt","M:\\CCC\\canterbury\\PLRABN12.txt");
yellow: StringHash_Test("m:\\ccc\\book1","m:\\ccc\\book1");
red:    StringHash_Test("M:\\CCC\\words\\Sentence-list_00,032,359_English_The_Holy_Bible.txt","M:\\CCC\\words\\Word-list_00,038,936_English_The Oxford Thesaurus, An A-Z Dictionary of Synonyms.wrd");
green:  StringHash_Test("m:\\ccc\\large\\enwik8","m:\\ccc\\book2");

make hash from words in first arg
lookup words in second arg

(generally second file is much smaller than first, so we get some hits and some misses)

```

Anyway, the result is that the Whiz and Jesteress variants of FNV1 by Georgi 'Sanmayce' are in fact quite good in this kind of usage. They rely on having fast unaligned dword reads, so they are pretty much Intel-only, which makes them a bit pointless. But here they are :

```
U32 HashFNV1(const char *key)
{
U32 hash = 2166136261;
while(*key)
hash = (16777619 * hash) ^ (*key++);
return hash;
}

inline uint32 FNV1A_HashStr_WHIZ(const char *str)
{
const uint32 PRIME = 709607;//15607;
uint32 hash32 = 2166136261;
const char *p = str;

for(;;)
{
uint32 dw = *(uint32 *)p;
if ( DWORD_HAS_ZERO_BYTE(dw) )
break;

hash32 = (hash32 ^ dw) * PRIME;
p += sizeof(uint32);
}
while( *p )
{
hash32 = (hash32 ^ *p) * PRIME;
p++;
}

return hash32;
}

U32 FNV1A_Hash_Jesteress(const char *str)
{
const U32 PRIME = 709607;
U32 hash32 = 2166136261;
const char *p = str;

for(;;)
{
uint32 dw1 = *(uint32 *)p;
if ( DWORD_HAS_ZERO_BYTE(dw1) )
break;

p += 4;
hash32 = hash32 ^ _lrotl(dw1,5);

uint32 dw2 = *(uint32 *)p;
if ( DWORD_HAS_ZERO_BYTE(dw2) )
{
// finish dw1 without dw2
hash32 *= PRIME;
break;
}

p += 4;

hash32 = (hash32 ^ dw2) * PRIME;
}

while( *p )
{
hash32 = (hash32 ^ *p) * PRIME;
p++;
}

return hash32;
}

```

I have no doubt these could be massaged to be a bit faster through careful study of the asm (in particular the handling of the 0-3 byte tail doesn't look awesome). An even better thing would be to ensure all your strings are 4-byte aligned and that after the null they have null bytes to fill the final dword.

Anyway, you can pretty much ignore all this, because hash functions have to be tested in-situ (eg. on your exact data, on your hash table implementation, on your platform), but it is what it is.

BTW FNV1 is a lot faster than FNV1a. Also the mixing of Whiz and Jesteress are quite poor and they do have a lot more collisions on some values, but that appears to not matter that much on this particular kind of test.

## 11/20/2010

### 11-20-10 - Function approximation by iteration , Part 3 - Reciprocal Square Root

See part 1 or part 2 .

Now that we have our toolbox, we can attack a popular issue in 3d engines - the reciprocal square root (used primarily in normalizing vectors). We want to find y = 1 / sqrt(x) .

Let me first state that I'm not doing all this because it's actually important for optimization. If you care about speed all you have to do is set :

```
/arch:SSE2 /fp:fast

```

To do reciprocal square root we're going to start with y0 from the IEEE floating point trick. In games this is commonly know as "Carmack's trick from Quake 3" , but it goes back to Greg Walsh apparently :

Also on the Slashdot page about that article Paul Hsieh mentions that apparently Kahan knew of this idea, but doesn't know the time frame exactly.

So the initial approximation is :

```
const int magic = 0x5f375a86;

inline float rsqrtf_walsh( float x )
{
union { int i; float f; } u;
u.f = x;
u.i = magic - (u.i>>1);
return u.f;
}

```
(more on the magic number later, but see Lomont for an explination of where it comes from). (BTW I don't mean to imply that it's "magic" by using that name; it's a standard IEEE bit trick).

This has rather large error so we are going to use this as y0 and find an iteration for y1.

We can now use our dimensional analysis method. Our unitless number must be

```
u = x * y0*y0

```
and our iteration must be :
```
y1 = y0 * f(u)

```
and we'll look at a few possible forms for f :
```
f1 = A + B * u

f2 = A + B / u

f3 = A + B * (1-u) + C * (1-u)^2

```
we can confirm that these forms are sane by finding them in other ways.

If we set f(y) = 1/y^2 - x , then the root f(y) = 0 occurs where y = 1/sqrt(x) , so we can use Newton's method to find the roots of f(y). The result is :

```
y1 = y0 * ( (3/2) - 1/2 x y0^2 )

or :

f1 , with A = 3/2 , B = - 1/2

```
If we set f(y) = y^2 - 1/x , the root is in the same place and we can use Newton's method and we get :
```
y1 = 0.5 * ( y0 + 1 / (x*y0) )

or :

f2 , with A = B = 1/2

```

If we use the first f(y) form and do a "Halley" or "Householder" iteration instead of a Newton iteration, that's form f3 with A = 1 , B = 1/2, C = 3/8.

```
(maximum % relative error) :

y0 : (no iteration) : 3.436526

y0 + one Newton     : 0.175126  (f1 with A = 3/2 , B = - 1/2)

y0 + two Newtons    : 0.000470

y0 + one Babylonian : 0.060595  (f2 with A = B = 1/2)

y0 + one Halley     : 0.010067  (f3 with A = 1 , B = 1/2, C = 3/8)

```
obviously these methods do not have equal computational complexity. In particular, the Babylonian method uses a divide, so it's much slower than the other methods. A serious study of function approximation should compare the quality to the CPU cost to make sure the method is at a good place on the speed/quality curve.

But of course we know the punch line is coming : we don't need to use those values for ABC. Because the initial estimate y0 has a known error range, we can optimize ABC over that range. We won't bother with doing it analytically, we'll just numerically search. The result is :

```
maximum percent relative error :

y0 : (no iteration) : 3.436526

y0 + one Newton     : 0.175126  (f1 with A = 3/2 , B = - 1/2)
method f1 optimal   : 0.087665  ( A = 1.501338 , B = -0.500461 )

y0 + one Babylonian : 0.060595  (f2 with A = B = 1/2)
method f2 optimal   : 0.030292  ( A = B = 0.499848 )

y0 + one Halley     : 0.010067  (f3 with A = 1 , B = 1/2, C = 3/8)
method f3 optimal   : 0.002523  ( A = 1.0 , B = 0.5011, C = 0.375608 )

```
by optimizing the coefficients we get roughly a 2X reduction of maximum error *for free* ! (4X for Halley - because it's quadratic; two newton steps would also get a 4X improvement). It's no more CPU work because a float constant multiply is a float constant multiply. We're just using better constants. As noted in the last post - these constants are no longer an iteration that you can repeat over and over, they are optimized for the first step only.

Lastly, let's go back to that magic number 0x5f375a86 that was used in the y0 approximation. Can we do better if we optimize that? The answer is basically no. (* see addendum)

```
In Quake3 Carmack used 0x5f3759df ; Carmack + 1 Newton has a max relative error of 0.175214 %

In his paper Lomont finds 0x5f375a86 is slightly better ; Lomont + 1 Newton gives 0.175129 %

```
but that level of tweakage is pretty irrelevant because they missed the big tweak, which is using an optimal iteration instead of a newton iteration :
```
inline float rsqrtf_cb_method1( float x )
{
const int magic = 0x5F37592F;
const float A = 1.501338f;
const float B = -0.500461f;

union { int i; float f; } u;
u.f = x;
u.i = magic - (u.i>>1);
float y0 = u.f;

float Bx = B * x;

return y0 * ( A + Bx * y0 * y0 );
}

rsqrtf_cb_method1 has a relative error of 0.087714 %

```
the magic number I happened to use is different than Lomont or Carmack, but it's actually irrelevant - you can use anything in 0x5f375000 - 0x5f376000 and you will get the same final error (but you will get slightly different values for A and B). The point is that the bias from the magic number can be balanced by optimizing A or B very effectively. So you don't have to worry about it - just get the magic number vaguely close. (* see addendum)

To demonstrate that this iteration should not be repeated, let's change :

```
return y0 * ( A + Bx * y0 * y0 );

to

float y1 = y0 * ( A + Bx * y0 * y0 );
return y1 * ( A2 + B2 * x * y1 * y1 );

```
now the errors for two steps are :
```
first step is cb_method1             : 0.087714 %

repeat same step ( A2 = A , B2 = B ) : 0.087716 %  (!!)

second step Newton (A2= 3/2, B2=-1/2): 0.000127 %

second step optimal                  : 0.000072 %  (A2 = 1.500000596f , B2 = -0.500000060f)

```

You should see obviously that the closer you get to the answer, the less optimizing helps, because the Newton iteration (which is based on Taylor series) becomes optimal as epsilon goes to zero.

Now before you get excited and go copy-pasting rsqrtf_cb_method1 around, let me stop you. As I said at the beginning, it's pointless because you have :

```
inline float rsqrtf_sse( float x )
{
return _mm_cvtss_f32( _mm_rsqrt_ss( _mm_set_ss( x ) ) );
}

with max relative error : 0.032585%

```
One Newton step after this takes you to 0.000029% error ; optimizing the iteration only gets you to 0.000026% so it's very pointless.

Jan Kadlec (rrrola) posted a comment pointing to his prior work on the topic . He finds a better initial magic number and then followup floats by doing a more rigorous search. What I wrote at (*) about the magic number not mattering wasn't quite right. I was doing greedy searches in the optimization, and got stuck in a local minimum. You have to have quite a good non-greedy searcher to find his solution. The exact value of it doesn't matter much, but there is this much better solution if you make a big jump to another neighborhood :

Copy-pasting rrrola's version for reference , with 0.0652 % max relative error.

```
float inv_sqrt_rrrola(float x)
{
union { float f; uint32 u; } y = {x};
y.u = 0x5F1FFFF9ul - (y.u >> 1);
return 0.703952253f * y.f * (2.38924456f - x * y.f * y.f);
}

```

Also, something you can perhaps help me with : Pizer had done even earlier work on this topic and roughly found the right (right = Kadlec) constant by finding the magic number that minimizes (1 + e_max)/(1 + e_min). Doing that he gets 0x5F1F1412.

If we say ratio = approx/correct , then what Pizer is minimizing is ratio_max/ratio_min , which is the same as minimizing log(ratio_max) - log(ratio_min) . I can't reproduce Pizer's result and I'm not sure why that would be the thing you want to minimize. Intuitively it does make sense that what you care about is really *log* relative error, not relative error. That is, going +100% is clearly better than going -100% in a relative sense (this is sort of like the fact that doing -100% in the stock market is much worse than doing +100% is good, as I wrote long ago, we should really talk about log-return).

I guess one way to see it is if you think about e as a floating point number; what we really want is for the error of the approximation to show up as far to the right of the decimal as possible. The number of binary digits to the right of the decimal is the log of the error. But that would just mean minimizing log( MAX(e_max,-e_min) ) so something is missing in my understanding of this bit.

### 11-20-10 - Function approximation by iteration , Part 2

See part 1 .

Let's get to an example. We want y = sqrt(x) , but without actually doing a square root.

You can get an iteration for this in various ad-hoc ways. Let's see one real quick :

```
y = sqrt(x)

square it

y^2 = x

divide by y

y = x/y

2y = y + x/y

y = 0.5*(y + x/y)

y_1 = 0.5*(y_0 + x/y_0)

```
and that's an iteration for sqrt. BTW in all these notes I'll be writing y_0 and y_1 as two steps, but I really mean y_n and y_n+1.

You can get this in another funny way :

```
y = sqrt(x)

y = sqrt( y * (x/y) )

y = geometric_mean( y , (x/y) )

y_1 = arithmetic_mean( y_0, (x/y_0) )

```
and you can see that it converges because if y is bigger than sqrt(x), then x/y is smaller than sqrt(x), so the mean must be closer to the answer.

But these are sort of ad-hoc weird ways of deriving the iteration. Let's do it in a more formal way which will be more useful in the general case. (BTW this is called the "Babylonian method for sqrt" and you can also get it trivially from Newton's method).

Say we have y_0 which is an approximation of sqrt(x) that we got some other way. It's close to the real answer y = sqrt(x) , so we write :

```
y = sqrt(x)

y^2 = y_0^2 * (1 + e)

or

y_0 = y / sqrt(1 + e)

e = "epsilon" is small

y_0^2 * (1 + e) = y^2 = x

1+e = x/y_0^2

e = x/y_0^2 - 1

```
the choice of how I've factored e out from (y/y_0) is pretty arbitrary, there are lots of choices about how you do that and all are valid. In this case the important thing is that e multiplied on y_0, not added (eg. I did not choose y^2 = y_0^2 + e).

So far this has all been exact. Now we can write :

```
y = sqrt(x) = sqrt( y_0^2 * (1 + e) ) = y_0 * sqrt( 1 + e )

exactly :

y = y_0 * sqrt( 1 + e)

approximately :

y ~= y_0 * approx( sqrt( 1 + e) )

so this is our iteration :

y_1 =  y_0 * approx( sqrt( 1 + e) )

```
The principle step in these kind of iterations is to take a twiddle-equals for y (an approximation of y) and make than an exact equals (assignment) for an iteration of y_n.

Now the only issue is how we approximate sqrt( 1 + e). First let's do it with a Taylor expansion :

```
approx( sqrt( 1 + e) ) = 1 + 0.5 * e + O(e^2)

y_1 = y_0 * ( 1 + 0.5 * e )

y_1 = y_0 * ( 1 + 0.5 * (x/y_0^2 - 1) )

y_1 = y_0 * ( 0.5 + 0.5 * x/y_0^2 )

y_1 = 0.5 * ( y_0 + x / y_0 )

```
which is the Babylonian method.

But as we discussed last time, we can do better. In particular, if we know the error bound of y_0, then we know the range of e, and we can use a better approximation of sqrt( 1 + e). Specifically if we use the IEEE float-int kind of method for finding y_0 :

```
float y_0(float x)
{
union { int i; float f; } u;
u.f = x;
const int ieee_sqrt_bias = (1<<29) - (1<<22) - 301120;
u.i = ieee_sqrt_bias + (u.i >> 1);
return u.f;
}

```
Then the maximum relative error of y_0 is 3.53% That means (1 - 0.0353) < |y_0/y| < (1 + 0.0353)
```
(1 - 0.0353) < 1/sqrt( 1 + e) < (1 + 0.0353)
(1 - 0.0353)^2 < 1/( 1 + e) < (1 + 0.0353)^2
(1 - 0.0353)^-2 > ( 1 + e) > (1 + 0.0353)^-2
(1 - 0.0353)^-2 -1 > e > (1 + 0.0353)^-2 -1
0.07452232 > e > -0.06703023

```
we'll expand the range and just say |e| <= 0.075 ; so instead of just using a Taylor expansion that assumes e is small, we can minimize the maximum error over that range.

We do that by expanding in the basis {1} and {x} (if you like this is the orthogonal Legendre or Tchebychev basis, but that's kind of silly to say when it's two functions). To find the optimal coefficient you just do the integral dot product with sqrt(1+e) over the appropriate range. The result is :

```
sqrt(1+e) ~= 0.999765  + 0.037516 * ( e/0.075 )

sqrt(1+e) ~= 0.999765  + 0.5002133 * e

y_1 = y_0 * ( 0.999765  + 0.5002133 * e )

y_1 = y_0 * ( 0.999765  + 0.5002133 * (x/y_0^2 - 1) )

y_1 = y_0 * ( 0.4995517  + 0.5002133 * x/y_0^2 )

y_1 = 0.4995517 * y_0 + 0.5002133 * x/y_0

```
We'll call this the "legendre" step for sqrt. Note that this is no longer an iteration that can be repeated, like a Newton iteration - it is specific to a known range of y_0. This should only be used for the first step after getting y_0.

And what is the result ?

```
max relative error :

y_0 ieee trick : 3.5276%

y_1 babylonian : 0.0601%

y_1 legendre   : 0.0389%

```
a nice improvement!

Now that we understand the principles behind how this works, we can just get a bit more hacky and find the answers more directly.

We want to find an iteration for sqrt(x) , and we have some y_0. Let's use a physics "dimensional analysis" kind of argument. We pretend x has units of "Widgets". Then sqrt(x) has units of Widgets^1/2 , as does y_1. We only have x and y_0 to work with, so the only unitless thing we can write is

```
u = y_0^2/x

```
(or any power of u), and our iteration must be of the form :
```
y_1 = y_0 * f(u)

for some function f

```
now , f can be any function of u, but to minimize the difficulty of the math we specifically choose
```
f(u) = A * (1/u) + B + C * u

```
now the issue is just finding the A,B,C which minimize the error.

The Babylonian/Newton iteration is just A = B = 0.5 (C = 0)

Numerical optimization tells me the optimal C is near 0, and optimal AB are near A = B = 0.499850

```
max relative error :

y_0 ieee trick : 3.5276%

y_1 babylonian : 0.0601%

y_1 legendre   : 0.0389%

y_1 numerical  : 0.0301%

```
the result is half the error of the babylonian method, for no more work - just a better choice of constant (0.499850 instead of 0.5).

## 11/19/2010

### 11-19-10 - Hashes and Cache Tables

Writing some things I've been perking on for a while...

There's a hash function test at strchr.com :

At first this test looks pretty cool, but on further examination it has a lot of flaws. For one thing, they're basically testing only english words. If what you care about is short english text, then maybe the results are good, but if you care about other things they are not. I do like the fact that they are testing the total time to compute the hash + the time to lookup. However, when you do that then the details of the hash table become crucial, and it seems they are using an externally-chained hash table which is a severe bias (I'd like to see tests using stlport std::hash_map, rdestl, khash, google's "dense hash", or the cblib hash_table, etc. - some standard reference hash implementations). You also need to talk about your hash fill ratio, whether you store the hash value, whether the data is interned, etc. These will affect which hash function is "best".

There are also major architectural issues. For example, the large multiplies of Murmur are not fast on all current platforms. Furthermore, on some platforms you can't do unaligned DWORD reads fast, and you may have to do endian swaps. These issues do not affect all hashes in the same way, so to compare hashes without considering these issues is a mistake.

Also, to do a fair comparison you really need to understand the details of each hash. Some are designed for use only in prime-size tables. Some should be collapsed by xor folding, others should not. You cannot simply plug a bunch of different hash functions into the same code and have a fair comparison.

But perhaps most crucially, the hashes are not optimized to the same amount. Some are unrolled, others are not. Some are reading dwords, some are reading bytes. This is not a fair comparison. Really the only way to do a perfect comparison is to optimize each hash function as much as possible, because hashes have a different degree of capacity for optimization. For example, Adler and Fletcher (which are checksums not hashes, BTW) can easily be made fully 16-byte parallel (see for example asmcommunity Adler32 ; there are also some very good Altivec versions floating around), most hashes cannot be directly parallelized the same way (they can only be "trivially" parallized in the sense of running multiple hashes at once in parallel on interleaved input streams).

(ADDENDUM : actually the biggest problem with the test is probably that the hash functions are called through function pointer, not inlined; this is a severe penalty for the cheaper hash functions; in particular with properly inlined hash functions I ran a similar test and found FNV to just destroy Murmur, which is the opposite of what they find)

Anyway, enough being a grouch, let's get to what I've been thinking about.

One piece of news : there's a (beta) update to Murmur2 called (unsurprisingly) Murmur3 :

Murmur3 improves the basic mixing a bit to fix some flaws in Murmur2. It also works on larger data elements, which is a mixed blessing. That will make throughput on large keys faster, but hurts performance on small keys. One of the nice things about Murmur2 was how simple it was.

And some links from the past for reference :

What I wrote previously about hash table lookup :

My final answer for hashing strings was a variant of FNV, hash value stored in the table, max 70% full, quadratic reprobing.

File hashing :

My final answer for file hashing was Bob Jenkin's Lookup3, modified to be 4-way SIMD. Since Lookup3 inherently works on 12 bytes at a time, the 4-way simd makes it work on 48 bytes at a time, and I then unroll that 8 times so that I can work on 384 bytes at a time (which is crucially 3 whole 128 byte cache lines).

So anyway. When you're talking about hashes it's important to distinguish what you're using them for. There are lots of variations, but I think there are three primary categories :

1. Hash for hash table lookup (with full resolution, either by in-table probing or external chains).

In this case, collisions are only bad in that they cause you to take more CPU time to probe. You will never fail to find what you are looking up due to collisions. There's a very subtle trade off here between hash quality and the time it takes to compute the hash - the winner will depend intricately on the exact data you are hashing and what kind of hash table implementation you are using. I'm sure this is one of those scenarios where different authors will all claim a different hash function is "best" because they were testing in different scenarios.

2. File checksum/integrity/corruption/hacking testing.

In this case you want the hash function to change any time the file has been changed. There are then two subsets of this : catching accidental corruption and catching intentional corruption. To check the ability to catch accidental corruption, you want to make sure your hash is robust to common types of damage, such as file truncation, injection of zeros, single bit flips, etc. Catching intentional corruption protects you from attacks. We're not talking about cryptographic security, but a good hash should make it very difficult to construct alternative data which is both *valid* and has the same hash.

3. Hashes for "cache tables". This is the main thing I want to talk about.

A "cache table" is a hash table in which collisions are not resolved (or they are resolved only some finite number of times). So far as I know there's not a standard term for this, but we found this Google Code project :

And I think "cache table" is a pretty good name for it. It's similar to CPU caches, but without a backing store. That is, a traditional "cache" is basically a "cache table", but then if the data is not found in the cache, you go and get it out of the slow location. We say "cache table" to make it clear that in this case there is no slow location - if it's not in the cache you don't find it.

Cache tables are very popular in data compression. To my knowledge they were first used popularly in LZRW. I then made heavy use of them for LZP (and PPMCB, which is very similar to the way modern compressors use them). They are now the basis of most modern compressors such as PAQ and related context mixers. Cache tables in traditional data compression usage never resize, you simply set your memory usage and that's how big the cache table is.

Usage of a cache table is something like this :

```
h = hash(key)

look up table[h]

(optionally) check if table[h] is right for key
if so return its data

optionally :
reprobe, either by stepping h and/or by searching "ways"

# of reprobe steps is finite
if failed to find, return not in table

```
There are a few optional places. A crucial one is whether you do exact resolution or simply accept collisions. (in the old language of LZP, this is "context confirmation"). That is, for minimum memory use, table[h] might not store anything about key. That means when hashes collide, you can't tell, and you might get back an entry which does not correspond to your key.

Sometimes this is just fine. For example, in LZRW style string matchers, if you get back the wrong data you will simply not get a string match. In other cases you might want some amount of collision resolution, but only probabilistically. For example, PAQ stores only a few bits of key in the table, so when hash collisions occur it can detect them most of the time, but accepts a certain amount of false positives. This issue should always be considered as a trade off of memory usage - is it better to use the memory to resolve collisions more accurately, or just to have a larger table so that fewer collision occur?

The other issue is how you reprobe, and that is closely related to your insertion strategy. When you try to insert and find a collision, you have a variety of choices. You can try all the possible reprobes for that hash value and pick the one that is "best" to replace under some criteria (perhaps such as the last used, or simply the slot that's empty). If you're using "ways" you might just do round-robin insertion in the ways, which is a probabilistic replacement strategy.

Anyway, I think "cache tables" is an interesting topic which I haven't seen explored much in the literature. It's particularly interesting in compression, because when you get a false collision (eg. you retreive data from the table and think it applies to you when it doesn't), that doesn't ruin your compressor, it just makes you code that event with more bits than ideal. So you can measure the cost of collisions as bit excess, and you ideally want to tune collisions to occur for less important data. Hash function collisions in cache tables affect both speed and the quality of results - eg. more collisions means more dropped data - but you also usually care exactly *which* data is dropped.

The question of reprobes within a linear tables vs. ways (multiple independent linear tables) is interesting, and sometimes the optimal solution is to use both. For example, my "medium speed" string matcher for the RAD LZ77 uses two independent hash values (in the same table) and 2 ways. That is, there are two linear tables, and you look at two spots within each table.

I think that Cuckoo hashing for cache tables is a particularly interesting possibility. "Cuckoo cache tables" could do the normal Cuckoo thing (where you have two hashes for each key and you swap data around to make space for an insertion), but you set a maximum number of swaps on insertion (something like 4 or 8), and if you have to do the cuckoo swap step and hit the maximum number, you just stop and let some item fall out of the cache. You can also easily extend this by keeping track of age or quality or something as you do the Cuckoo, and choose to drop the worst item instead of simply dropping the last one in the chain. You don't have to worry about the pathological Cuckoo case of having a cycle you can't resolve - you simply drop an item in that case (in fact you don't even have to detect that case, you'll just go around the cycle up to your maximum finite number of steps).

(ADDENDUM : I tried cuckoo cache tables )

### 11-19-10 - Games without Strings

Quite a number of major video games have shipped with embarassingly large amounts of time spent in "strcmp" and embarassingly large amounts of memory spent on strings.

I'm of the opinion that games should not ship with strings. However, strings are very useful for development. How should we reconcile this? The basic strategy in all cases will be to replace the string with just an integer GUID or index. Then instead of strcmp you just do == test on the ints, and instead of storing lots of strings in memory, you just have ints to refer to other objects.

First of all, let's say that the premature optimization of making all your references into compact indices from the beginning is a mistake IMO. The main place that strings come up and where they are super useful is as the names of content. That is, textures, models, prefs, NPC types, weapon types, levels, attributes, etc. If I want to make an NPC it's nice to just be able to say, okay use the "ToughMarine" model and give him "AggressiveEnemy" behavior and "BigShotgun" weapon. Often those strings correspond to file names, either directly or indirectly, which is very handy because it lets you add new elements by just making new files, and it lets you find the content that a string corresponds to.

The bad premature optimization that I'm talking about is to work with "databases" of content right from the start rather than loose files. Then your references can be indices. So instead my NPC would use model 7, behavior 3, and weapon 41. This makes your references fast and compact right from the start, but the penalties to development process are just too severe. (some of the old game engines did this, but I believe it is now understood in the industry that this is very bad for source control and large teams and is being phased out).

So, we don't want that method, we want to work with strings during development, but ship without strings.

( some example of unnacceptable workflow in my opinion : when artists have to yell out "who has the texture name database checked out? I have to add a new string to it" , or when there's an error in the content hookup and the game prints out "error : npc 13 couldn't find model 9" (though, kudos for at least printing an error) )

One option is to have a "bake" step that converts all the loose content and string references into these packed databases and indexes. Basically it has to load up every piece of content, see what all the strings in the game are and convert them to an index, then order all the data by its index so you can find it. While this does work, it's pretty painful. Whole-game bake operations like this are pretty disastrous for development, because it means you can't test the real shipping game without doing some big time consuming process. I'm a big believer in being able to run at least portions of the game in their real shipping state all the time during development, not just at the very end. It makes it hard to just change one piece of content or add something and have.

Another option is to have a sort of "incremental baking" from a GUID server. I've seen this work in practice, so it is viable, but it's a little scary. Basically the idea is you keep a central string table that maps unique strings to indices (and vice versa). Each time you make a new piece of content or make a new string, you have to send a packet to the GUID server and get the index for that string. Now you can refer to your piece of content by that index, so you have compact references. While this does certainly work, relying on being able to communicate with the GUID server for development is a bit scary. Also if you ever accidentally get a bug in the GUID system you could corrupt a bunch of content. To deal with both of those issues, it would be nice to keep the original strings around in the content files during development as backup for the central repository.

The option we used for Oddworld Stranger was string hashing. In the shipping game, every 32 bit char * was replaced with a 32 bit integer hash of the char *. Using a hash makes the string -> index mapping deterministic and local, you don't have to talk to your neighbor to make sure you are getting a unique int. This method has lots of advantages and only one (rather large) disadvantage : the possibility of hash collisions. When you get a hash collision you wind up with a perplexing but amusing bug, like you try to put "SinisterHelmet" on a guy and instead you get "RubberChicken" attached to his head because those string hashes collided. During development you need to keep both the hash and the string around.

To handle collisions, we had a debug mode that would load up a level with the strings and the hashes and check if any strings had the same hashes. Note that you don't need to check collisions across the whole game, but only for content that can possibly be loaded at the same time. On Stranger I think we wound up with 2 collisions over the whole course of development. The solution to those collisions was simply to rename some content. eg. "FloppyHat" was colliding with "PurpleSparkle" , so we just renamed it to "PurpleSparkle2" and we were hash collision free. I definitely do not advocate the 32-bit hash method for large scale game development. It's okay on a small team where you can check things like that once in a while and handle it, but with big distributed teams and huge amounts of content it's not viable.

The simplest fix is to go to a 64-bit hash. Then the probability of collision becomes so small that you could just deal with it in the unlikely event that it happens. Note that in the shipping game you never actually generate these hashes, so the speed of the hash function is irrelevant; in the final game they are opaque GUIDs that are used to link pieces of content together.

With the hash+string method there's also an issue of how you store your content and what you do with the string in the shipping version of the game. What we did is just to store both inline in the data files (eg. 4 bytes of hash then the string bytes immediately following), and in the shipping game you just read the string in then immediately discard it. But if you want to be really super optimal and flat-load your data, then you would have to factor out the strings at some point. One option is to make all your data files in chunks, and make one of the chunks be "nonfinal" data, and put the hash -> string mapping in the nonfinal data chunk. Then you can "flat load" the rest of the data.

But none of these solutions is definitively awesome, so I'm curious what other people do or if there is a better solution.

Mick West wrote about it for game developer but doesn't really address the difficult issues.

Ivo Beltchev describes CRCView - a VC extension to do GetStr() for CRC's. I wonder how well that works; I've always found the VC expression evaluator to stop working right when I need it most.

Mischief.mayhem puts it together. He suggests using an external SQL database for GUID->String mapping and collision resolution.

BTW I'm not sure why you bother with hashing if you are using a global database. Just make the first string be "1", the second string be "2", etc. But I'm not super convinced that the global database is an okay method.

## 11/18/2010

### 11-18-10 - Bleh and TID2008

Bleh, I just spent the whole day trying to track down a bug in my MS-SSIM implementation, that turned out to be a total red herring. I was trying to test my MS-SSIM on the TID2008 database to confirm their results (they find that MS-SSIM is the best by far). (note that perceptual image papers currently have the nasty property that every author demonstrates that their method is the "best" , because they all use different testing methods and on different databases).

Anyway, I was getting totally wrong results, so I went over my MS-SSIM implementation with a fine toothed comb, checked everything against the original Matlab; I found a few deviations, but they were all irrelevant. The problem was I was running the files in the TID database in a different order than the TID test program wanted, so it was matching the wrong file to the wrong file.

As part of that, I downloaded all the implementations of MS-SSIM I could find. The best one is in Metrix Mux . Most of the others have some deviation from the original. For example, many people get the Gaussian window wrong (A Gaussian with sdev 1.5 is e^(- 0.5 * (x/1.5)^2 ) - people leave off the 0.5), others incorrectly apply the window at every pixel position (you should only apply it where the whole window is inside the image, not off any edge), another common flaw is to get the downsample wrong; the way authors do it is with a Daub9 lowpass tap and then *point* downsampling (they use :1:2:end matlab notation which is a point downsample). Anyway, a lot of these details are pretty random. Also of note : Metrix Mux uses rec601 luma and does not gamma correct.

The TID2008 perceptual distortion database is a lot better than the Live database, but it's still not great.

The problem with both of them is that the types of distortions applied is just too small of a subset. Both of them mainly just apply some random noise, and then they both apply JPEG and JPEG2000 distortions.

That's okay if you want a metric that is good at specifically judging the human evaluation of those types of distortions. But it's a big problem in general. It means that metrics which consider other factors are not given credit for their considerations.

For example, TID2008 contains no hue rotations, or images that have constant luma channels but visible detail in chroma. That means that metrics which only evaluate luma fidelity do quite well on TID2008. It has no images where just R or just G or just OpponentBY is modified, so you can't tell anything about the importance of different color channels to perception of error.

TID2008 has 25 images, which is too many really; you only need about 8. Too many of the TID images are the same, in the sense that they are photographs of natural scenes. You only need 2 or 3 of those to be a reasonable representative sample, since natural scene photos have amazingly consistent characteristics. What is needed are more distortion types.

Furthermore, TID has a bunch of distortion types that I believe are bogus; in particular all the "exotic" distortions, such as injecting huge solid color rectangles into the image, or changing the mean luminance. The vast majority of metrics do not handle this kind of distortion well, and TID scores unfairly penalize those metrics. The reason it's bogus is that I believe these types of distortions are irrelevant to what we are doing most of the time, which is measuring compressor artifacts. No compressor will ever make distortions like that.

And also on that thread, too many of the distortions are very large. Again many metrics only work well near the threshold of detection (that is, the distorted image almost looks the same as the original). That limitted function is actually okay, because that's the area that we really care about. The most interesting area of work is near threshold, because that is where we want to be making our lossless compressed data - you want it to be as small as possible, but still pretty close to visually unchanged. By having very huge distortions in your database, you give too many points to metrics that handle those huge distortions, and you penalize

Lastly, because the databases are all too small, any effort to tune to the databases is highly dubious. For example you could easily do something like the Netflix prize winners where you create an ensemble of experts - various image metrics that you combine by estimating how good each metric will be for the current query. But training that on these databases would surely just give you overfitting and not a good general purpose metric. (as a simpler example, MS-SSIM has tons of tweaky hacky bits, and I could easily optimize those to improve the scores on the databases, but that would be highly questionable).

Anyway, I think rather than checking scores on databases, there are two other ways to test metrics that are better :

1. Metric competition as in "Maximum differentiation (MAD) competition: A methodology for comparing computational models of perceptual quantities". Unfortunately for any kind of serious metrics, the gradient is hard to solve analytically, so it requires numerical optimization and this can be very time consuming. The basic method I would take is to take an image, try a few random steps, measure how they affect metrics M1 and M2, then form a linear combo of steps such that the affect on M2 is zero, while the affect on M1 is maximized. Obviously the metrics are not actually linear, but in the limit of tiny steps they are.

2. Metric RD optimization. Using a very flexible hypothetical image coder (*), do R/D optimization using the metric you wish to test. A very general brute for RD optimizer will dig out any oddities in the metric. You can test metrics against each other by making the image that this compressor will general for 1.0 bits per pixel (or whatever).

* = you shouldn't just use JPEG or something here, because then you are only exploring the space of how the metric rates DCT truncation errors. You want an image compressor that can make lots of weird error shapes, so you can see how the metric rates them. It does not have to be an actual competitive image coder, in fact I conjecture that for this use it would be best to have something that is *too* general. For example one idea is to use an over-complete basis transform coder, let it send coefficients for 64 DCT shapes, and also 64 Haar shapes and maybe some straight edges and linear ramps, etc. H264-Intra at least has the intra predictors, but maybe also in-frame motion vectors would add some more useful freedom.

## 11/16/2010

### 11-16-10 - A review of some perceptual metrics

In general this appears to be an active area of research, with tons of good work coming out in just the last few years. It's finally getting the attention it deserves, thank god.

In general there are two major approaches to perceptual metrics : HVS masking/filtering methods , and structural/statistical methods.

The HVS methods are mainly still using something like a per pixel delta, but they are weighting or masking or remapping it to account for various perceptual effects. Largely they are considering the capabilities of the eye, as opposed to what the brain processes and takes from the image. This includes stuff like rod/cone spatial resolution, perceptually linear response to intensity differences, contrast masking, luma masking, relative contrast sensitivity, etc.

The original major HVS work was the JPEG standard itself with its quantization matrices that are based on CSF (contrast sensitivity function) measurements. After that Watson's DCTune and related work added masking and other effects. This has been extended over the years. Probably the strongest modern work is Lin's JND (just noticeable differences) stuff.

The JND papers work out complicated models of visual masking to figure out what kind of deltas between images are imperceptible.

Obviously color spaces like CIELAB etc are part of the HVS image difference metric, as is the SCIELAB filter. (I guess CIEDE2000 is the newer version)

A lot of the HVS stuff doesn't really deal well with color - most of the old data is for luma only. The Fairchild/iCAM group is doing a lot of work on perception of color and mainly HVS based metrics :

Garrett Johnson's Thesis is the best reference I have found for lots of data on the HVS with actual useful formulas and numbers.

Many of the newer works use a whole-image FFT in opponent color space so that they can apply a CSF (ala JPEG quantization matrices) and a rod/cone spatial resolution filter (ala SCIELAB).

There are lots of metrics made over the years that incorporate some amount of these HVS factors, for example PSNR-HVS-M :

For video there are more HVS factors to consider such as temporal contrast sensitivity and motion masking.

There are also a lot of papers on contrast visibility and color appearance models from the HDR Tone Mapping research guys; in fact there are some good Siggraph courses such as Image Quality Metrics from Siggraph 2000 and Survey of Color for Computer Graphics . There were also a bunch of papers on selective raytracing that gave good details on perceptual metrics.

However, in general the HVS-centric methods are a bit out of favor and "structural methods" appear to be better. Again most of the structural methods work on luma only and ignore color.

The most well-known structure method is SSIM , and there have been many extensions of SSIM that purport to improve it, such as gradient-SSIM and region-pooled-SSIM or feature-type (edge,smooth,texture) pooled SSIM.

The LIVE team has a ton of good papers, with the best ones being on "structural" type metrics :

In particular, they have VIF (Visual Information Fidelity) which is a measure of human-perceptual mutual information. There are a number of modern works using metrics like this for image quality which appears to be a very promising line.

There are various other structural-type metrics, such as doing SVD to find singular vectors and values and so on. So far this work feels to early to be useful.

One interesting work on color I found is SHAME : (Spatial Hue Angle Metric) :

It's basically SCIELAB with the addition of a hue histogram. The hue histogram weights colors that occur more often more. eg. if your image has a big patch of red and you get the color wrong on it, that's visible, but if it only has a few scattered red pixels, then their exact color doesn't matter because relative to their surrounding they will seem "red" even if they are pretty far off.

Some other random links I thought were interesting :

## 11/12/2010

### 11-12-10 - Some notes on function approximation by iteration

People often blindly use Newton-Raphson method to iteratively improve function approximations. It's worth taking a moment to think about where that comes from and how we can do better.

First of all, iterative function approximation almost always comes from some kind of epsilon expansion (often Taylor series). You assume you are near the desired answer, you set the discrepancy to epsilon, then expand and drop higher powers of epsilon.

For example, Newton-Raphson iteration is this :

```
You wish to find a root r of f , eg. f(r) = 0

Assume I have x_0 which is close to r, eg, r= x_n + e

0 = f(r) = f(x_n + e) = f(x_n) + e * f'(x_n) + O(e^2)

so

e = -f(x_n)/f'(x_n) + O(e^2)

r = x_n - f(x_n)/f'(x_n) + O(e^2)

therefore

x_n+1 = x_n - f(x_n)/f'(x_n)

is closer to r than x_n

```
The way we use this to do function approximation is we set up a function f() such that the root of f is at the function we wish to approximate.

For example to find sqrt(A) you would set f(x) = x^2 - A ; this f has roots where x^2 = A , eg. x = sqrt(A).

Note that this method is only really useful when you have a function like sqrt where the function is expensive but its inverse is cheap. This is where the "art" comes in - there are any number of functions you can write which have the root at sqrt(A), but you need to write one that doesn't involve a sqrt; the whole point is to find an iteration that is much cheaper than the full quality function. For example to approximate reciproal sqrt you could also use f(x) = x^2 - 1/A or f(x) = 1/x^2 - A , both are valid but produce different iterations of different complexities.

So, now that we sort of see where Newton-Raphson iteration comes from, we can throw it away. What we really want is a more general idea - start with a point close to the solution, do an epsilon expansion, solve for epsilon, that gives us an iteration. In particular, Newton-Raphson is doing an expansion of the form r = x_n + e , but we could also do r = x_n * (1+e) or r = x_n/(1+e) or whatever. Sometimes these give better iterations (in terms of complexity vs. accuracy).

The next major general point to remember is that all of these methods are based on something like Taylor expansion, and while Taylor expansion is great for infinite series, we should all know that it is *not* optimal for finite series. That is, the first N terms of the Taylor expansion of a function are *not* the best N-term polynomial expansion of that function (except in the limit as epsilon goes to zero).

Over a finite interval, you could use Legendre Polynomials to find the actual best N-term approximation (or any other orthogonal polynomial basis). Since we are on computers here it's often easier to just do a brute force solve for the best coefficients.

This is well known to people who do function approximation (eg. any sane person doing a polynomial approximation of sin(x) over the range [0,pi] is not using a Taylor expansion) (see previous discussion of this 06-13-09 and 06-21-09 ). However, when people do iterative approximation this idea goes out the window for some reason.

In particular, after your initial approximation, you usually know something about the error, whether it is positive, what the bound is. Then you can expand in the error and find the optimal coefficients over that range. This gives you an iterative refinement step which is superior to Newton.

I'll do a followup post and work a specific example.

## 11/09/2010

### 11-09-10 - New Technology Fucking Blows

I recently "upgraded" to an HDTV and a PS3 and have been watching some Blu Rays.

WTF. Why didn't you all tell me this is a fucking disaster ?

First of all, the fucking Blu Rays want to connect to the internet. Umm, no you don't. Second, they want to run fucking Java to make their menus even more fucking annoying than the damn DVD menus.

Okay, friggle frack, I hate it, but I can live with that.

Then I find YOU CAN'T FUCKING RESUME on BD-J discs !? WTF "Resume" is like the most basic feature that you need in a movie player, like I'm not allowed to fucking stop my movie and then start it again !? I love the "recommended fixes" from even the official Sony page, to just "skip ahead to where you stopped watching". Are you fucking kidding me? Somebody should be shot for this, it's like the engineers making a video format without audio. You add a bunch of menu features and shit and oh, by the way it doesn't play audio anymore. You can't break the basic fucking functionality when you add features!

Okay, so all Blu Ray players have that problem, but the PS3 is actually even worse than most and won't even Resume from normal blu rays unless you fucking say a hail mary and cover it with garlic and press just the right buttons. And god forbid it do something like remember the location you stopped watching even after you power it down or remove the disc, like you know every fucking DVD player from 10 years ago does. ( PS3 missing resume function ).

(aside : that PS3 link is a pretty funny example of people defending their enemies :
"damn just remember the chapter before shutting down."
"quit being lazy and find the chapter where you left off and then press play on that"
"I just leave mine on 24/7 and if I stop a movie, it resumes where I left off until I eject it."
WTF R u serious? )

And of course every time you do watch a Blu Ray, you get the wonderful fucking joy of sitting through a bunch of mandatory fucking previews and other nonsense because they "forbid that operation". Yeah yeah you can usually just skip chapters to get to the menu, but you can't go direct to top menu on most discs, and you can't jump directly to playing.

As for HDTV, it's mostly good. The fucking Bezel on my LG is glossier than greased shit, which is pretty damn awful. Also HDMI connectors are the fucking devil, they're like bad old USB connectors, they don't go in far enough and are wobbly. I have problems with the HDMI connector making good stable contact at the TV and at the PS3 and the computer.

Oh, and of course neither the TV nor the PS3 came with cables even though they're like fifty cents OEM price now. That's just fucking cheap ass annoying shit.

So far the only really huge win is that the Netflix app for the PS3 is way better than the one for the HTPC, and of course Netflix movies fucking start playing directly without a bunch of shit, and you can resume them. I would just use Netflix 100% of the time except for the fact that Brownstripe is not reliable, and there's nothing more annoying than being in the middle of a movie and your net connection decides to flake out (I wish I could fucking locally cache the whole movie before watching it to ensure full playback).