Presentation is loading. Please wait.

Presentation is loading. Please wait.

STEREO MATCHING USING POPULATION-BASED MCMC

Similar presentations


Presentation on theme: "STEREO MATCHING USING POPULATION-BASED MCMC"— Presentation transcript:

1 STEREO MATCHING USING POPULATION-BASED MCMC
Joonyoung Park1 Wonsik Kim2 Kyoung Mu Lee2 1LG Electronics Inc. 2Computer Vision Lab. School of EECS Seoul National University, Korea Good afternoon. My name is Joonyoung Park. This work was co-worked by Wonsik Kim and my professor Kyoung Mu Lee. I will talk about stereo matching using population-based MCMC, which is one of the sampling-based methods.

2 Low-Level Vision as an Optimization Problem
Stereo Photomontage Segmentation Denoising /Inpainting . In various computer vision problems like stereo, Photomontage, Segmentation, Denoising, and Inpainting, In order to solve them, we design some energy models like this form, Then we can change the problems into finding a solution that minimizes Energy. That is an optimization problem. So optimization is one of the most important part in the computer vision area. General pairwise MRF model:

3 Existing Methods for Optimization Problem
fast easy to get stuck at local minimum Graph cuts α–Expansion / Swap Message passing Belief Propagation (BP) / Tree-ReWeighted Message Passing (TRW) Sampling based method Gibbs sampler Swendsen-Wang Cuts There have been many tries to solve the optimization problem, And recently, there are two kinds of methods which are widely used. They are Graph cuts based method and message passing based method. But they have a risk to get stuck in local minima because they are deterministic methods, On the other hand, We have an alternative, which is sampling based method. Because it is statistical method, it is able to find global minimum. But there is a big problem in sampling based approach. It’s too slow. So we use some acceleration techniques, giving up the guarantee of global minimum. slow global minimum

4 Our Study Population based MCMC We develop a new efficient
sampling based optimization method. Population based MCMC In our work, we propose to use a new kind of sampling-based method, Population-based MCMC. From now, I will call it Pop-MCMC We adopt this method from statistics, and develop it for vision problems in a proper way.

5 What is sampling ? I’ll start with the concept of sampling.

6 Sampling To generate samples from a target distribution
The original goal of sampling is to generate samples from a given distribution. For example, the distribution is given like this, and we extract samples like this figure by sampling. Given probability distribution Obtained samples

7 Optimization Two important concepts Annealing + Sample move
previous sampling based methods Two important concepts Annealing + Sample move But here, we are trying to use sampling in the optimization problem. For optimization, there are two important concepts. One is annealing and the other is sample move. I will give you more explanation about these concepts in detail.

8 ... ... Optimization target distribution: temperature:
previous sampling based methods Annealing target distribution: temperature: ... The First part is annealing. The target distribution is given by pi, and we introduce a new variable for temperature T. At first we start from high temperature, and as time goes by, the temperature decreases. At high temperature, the distribution is almost flat, and at low temperature, the distribution becomes concentrated on global optimum. ...

9 ... ... Optimization target distribution: temperature:
previous sampling based methods Annealing target distribution: temperature: ... We perform the sampling from high to low temperature, At high temperature, the samples are wandering freely, But as the temperature becomes low, the samples become concentrated on the global optimum. ...

10 ... ... Optimization target distribution: temperature: Most Probable
previous sampling based methods Annealing target distribution: temperature: ... Most Probable State! So finally at the lowest temperature, we obtain the global optimum. ...

11 Optimization Previous methods Gibbs sampler Swendsen–Wang Cuts
previous sampling based methods Sample move Previous methods Gibbs sampler Swendsen–Wang Cuts Then, how can we obtain the samples? We usually get the next sample from the current sample, and we call it sample move. I’ll talk about previous methods of sample move.

12 Optimization Flip the label of single node Too slow (only local move)
previous sampling based methods Sample move Prev. Method①: Gibbs Sampler The left one represents the current sample, And the right one represents the next sample. Gibbs sampler is the most primitive method. In Gibbs sampler, we can flip the label of only one node, So it takes very long time to converge to global optimum. Flip the label of single node Too slow (only local move)

13 Optimization Flip the label of a cluster
previous sampling based methods Sample move Prev. Method②: Swendsen-Wang Cuts And in Swendsen-Wang cuts, we can flip the labels of multiple nodes at the same time. At first, we make a cluster with some probabilistic rule. And then we flip the label of the cluster. Here, All the nodes of one cluster should have same label, it is one of the limitation in SWC. Anyway The move becomes rather bigger than Gibbs sampler, but it is only local move, so it is still slow Flip the label of a cluster Still too slow (only local move)

14 Limitation of previous methods
Optimization previous sampling based methods Limitation of previous methods slow convergence rate Fast cooling: easy to get stuck at local minima As I said, the limitation of these methods is slow convergence rate. So in practical case, we use fast cooling schedule in the process of annealing. And then, it is easy to get stuck in local optima, not in global optimum.

15 ... ... Optimization Parallel tempering Conventional annealing:
time Parallel tempering: time time To overcome this problem, parallel tempering was proposed. Unlike conventional annealing, it uses multiple chains simultaneously. time ... time

16 ... ... Optimization Parallel tempering Conventional annealing:
time Parallel tempering: time interaction! time And there occur some interactions between chains. It makes global move, Therefore we can make convergence rate faster. But it is not enough, so Pop-MCMC is proposed. interaction! time ... interaction! time

17 Population based MCMC It was originated from parallel tempering
It uses multiple chains It allows more active interactions between chains. We designed 3 sample moves Mutation : The move of single chain Exchange : Crossover : Ways of interaction between two chains Now, we are ready to talk about Pop-MCMC. It was originated from parallel tempering, so it also uses multiple chains. But it allows more active interations between chains. And it uses three kinds of move. mutation, exchange, crossover are those. And the new design of mutation and crossover moves are our contribution. I will talk about these moves in more detail. Our contribution!!

18 PopMCMC: Move① – Mutation
Optimization Population based MCMC PopMCMC: Move① – Mutation The first move is mutation.

19 ... Optimization Population based MCMC PopMCMC: Move① – Mutation time
In mutation move, At first, one chain is chosen randomly among the multiple chains. ... time

20 ... Optimization Population based MCMC PopMCMC: Move① – Mutation time
And for the sample of chosen chain, we perform a sample move like conventional sampling methods. Here, SWC which I said before is applied. ... time

21 PopMCMC: Move② – Exchange
Optimization Population based MCMC PopMCMC: Move② – Exchange The second move is exchange.

22 ... Optimization Population based MCMC PopMCMC: Move② – Exchange time
In exchange move, Two chains are randomly chosen. ... time

23 ... Optimization Population based MCMC PopMCMC: Move② – Exchange
time time time And for the samples of two chosen chains, we exchange two samples. ... time

24 PopMCMC: Move③ – Crossover
Optimization Population based MCMC PopMCMC: Move③ – Crossover The third move is crossover.

25 ... Optimization Population based MCMC PopMCMC: Move③ – Crossover time
In crossover move, two chains are randomly chosen like the case of exchange move. ... time

26 ... Optimization Population based MCMC PopMCMC: Move③ – Crossover time
And for the samples of the chosen chains, we decide a cluster to exchange. In SWC, The cluster must be composed of same labels. But in this crossover move, there is no limitations in making cluster. I’ll skip the mathematical part which makes this possible. If you want it, you can refer the paper. We can make a cluster just randomly, and it enables global move. After making cluster, the crossover moves are performed by exchanging two parts like this. ... time

27 ... PopMCMC: Overall mutation mutation mutation mutation mutation
crossover /exchange crossover /exchange mutation mutation mutation mutation mutation crossover /exchange crossover /exchange mutation mutation mutation mutation mutation This is the overall mechanism of popmcmc. For each chain, the mutation moves occur, and between randomly chosen two chains, the crossover or exchange moves occur. ... crossover /exchange crossover /exchange mutation mutation mutation mutation mutation

28 Proposed: Population-based MCMC
Advantages Global move Unlike Swendsen-Wang Cuts that performs only local move, our exchange and crossover moves allow global move. Fast Mixing Global move and parallel tempering enable fast mixing. The advantage of Pop-MCMC are like this. Unlike previous methods, we can perform global moves by exchange and crossover moves, And the global move makes mixing rates faster, it means that we can approach global optimum faster.

29 using Segment-based Energy Model
Application to Stereo Stereo Problem using Segment-based Energy Model (A. Klaus et. al., ICPR 2006) We apply Pop-MCMC to stereo problem. The used Energy model is adopted from Klaus’ paper in ICPR 2006

30 Segment-based Energy Model
Application to Stereo Segment-based Energy Model At first we over-segmented the reference image using mean-shift algorithm, And then we consider each segment as a node, and adjacent segments as neighbors. Each node will have a label, which is a plane of disparities. So the energy function is given like this. = MRF DESIGN = NODES: segments NEIGHBORS: adjacent segments LABELS: planes of disparities

31 From Middlebury web-site
Application to Stereo Test Images The test images were taken from middlebury website, Tsukuba Venus Teddy Cones From Middlebury web-site

32 Experimental Results Disparity maps Tsukuba Venus Teddy Cones 1.38
Bad pixels(%) 1.38 1.21 14.7 13.1 And these are the results of disparity maps. In fact, the rate of bad pixels are not so low comparing with the state of the art. It is because we just use the simple energy model.

33 Experimental Results Convergence Tsukuba Cones
However what is important is these graphs. In the case of using the same energy model, Comparing with the other methods like Belief Propagation or SW-Cut, We can find lower energy state by using Pop-MCMC.

34 Conclusion Future works
We develop a new efficient Population-based MCMC for optimization and applied it to stereo problem. It finds lower energy state compared with other methods. Other applications / More complex energy models Future works I’ll summarize my talk. In our work, we develop a new efficient Pop-MCMC and applied it to the stereo problem. And it finds lower energy sate compared with other methods. So it shows a possibility to be used as an optimization tool. For future work, we can apply it to other application, and we can use the more complex energy models for improving the quality of disparity maps

35 Seoul National University
Thank You Computer Vision Lab. Seoul National University Thank you for your attention.


Download ppt "STEREO MATCHING USING POPULATION-BASED MCMC"

Similar presentations


Ads by Google