Stereo Vision Laboratory

4. Comparison of parallax images

Paper comparing various algorithms

The Middlebury Computer Vision Pages 's MRF Minimization page, well-known for computer vision benchmarks introduced in Article 2, includes `` A Comparative Study of Energy Minimization Methods for Markov Random Fields with Smoothness-Based Priors, IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 30 (6): 1068-1080, June 2008 "is introduced. This paper is a detailed comparison of the energy minimization algorithms for the site labeling problem. It accurately describes what energy function was used and the number of labels.

And on the page of Result Tsukuba stereo , parallax images that are the results of algorithm execution for Tsukuba stereo images are introduced for each algorithm. The energy function used to minimize the energy of this Tsukuba stereo image seems to be as follows regardless of the algorithm. First, the data term is the sum of the absolute values of the differences for each RGB, which is exactly the same as Article 3. Next, the smooth term is truncated linear, with a difference censored at 2, and the coefficient is 40 if the pixel value change in the left image is less than 8, and 20 otherwise. This is much more complicated than article 3. It seems that you are trying to handle the troublesome label of parallax as well as possible. The number of labels is 16. The pixel-disparity model in article 3 uses 19 labels from parallax-10 to parallax-28, and the gaze_line-depth model uses 10 labels from depth-5 to depth-14 because depth is half the parallax. On the other hand, the compared algorithm and the resulting disparity file is

List of algorithms
abbreviation	description	disparity_file
ICM	Iterated Conditional Modes	tsu-ICM.png
Swap	graph cut, Swap move	tsu-Swap.png
Exp	graph cut, Expansion move	tsu-Expansion.png
BP-S	Belief Propagation, S version	tsu-BP-S.png
BP-M	Belief Propagation, M version	tsu-BP-M.png
TRW-S	Tree ReWeighted message passing, S version	tsu-TRW-S.png
GT	Ground Truth	tsukuba-truedispL.png

. I have downloaded these parallax files. Please also see disparity14.png and pen_14_inh_1023.png created in Article 3 and true.png created in Article 2. In the parallax images that have been downloaded, the right side of the orange standlight is clear.

A program that simply compares parallax

I created a program compare_disparity.cpp that simply compares the parallax between two parallax files. This program compares the disparity of two disparity files for each pixel and sums the differences. However, when the parallax of the standard parallax file is 0, the parallax difference of that pixel is ignored. makefile If you also download

make do

, the difference from the above parallax file will be output based on true.png created in Article 2. As you can see from the makefile, the program starts with

./compare true-disparity_file disparity_file

Comparison result

The comparison result with the modified true parallax image true.png created in Article 2 is

Comparison with modified true disparity
abbreviation	description	disparity_file	difference
ICM	Iterated Conditional Modes	tsu-ICM.png	3118976
BP-S	Belief Propagation, S version	tsu-BP-S.png	691440
BP-M	Belief Propagation, M version	tsu-BP-M.png	430864
Expansion	graph cut, Expansion move	tsu-Expansion.png	338960
Swap	graph cut, Swap move	tsu-Swap.png	334416
TRW-S	Tree ReWeighted message passing, S version	tsu-TRW-S.png	329920
DIS	pixel-DISparity model	disparity14.png	249344
DEP	gaze_line-DEPth model	pen_14_inh_1023.png	216336
GT	Ground Truth	tsukuba-truedispL.png	0

. If you make this a stick rough, it will look like

Comparison with modified true parallax image

. ICM was too big to put in the bar graph. Looking at this, the result of gaze_line-depth model seems to be very good DEP. The gaze_line-depth model has only half the resolution in the depth direction. The result of minimizing the energy of the pixel-disparity model with a linear energy function is not bad. Possible reasons for this are some mistakes in this series of articles, it is not good to simply compare the parallax of pixels, otherwise it is approximate except for DIS and DEP, and the annoying nature of parallax I wonder if the energy function devised for the purpose was not useful.