1

NEAR REAL-TIME STEREO MATCHING USING GEODESIC DIFFUSION

IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, VOL. 34, NO. 2, FEBRUARY

2012

Leonardo De-Maeztu,Arantxa Villanueva, Member, IEEE, and

Rafael Cabeza

Guan-Yu Liu

2

Outline Introduction Overview Related work Method Experimental Results CUDA Q & A

3

Introduction(1/4)

Stereo matching Local matching

A finite region(window size) is being computed Global matching

Do smoothness by energy minimization techniques

4

Introduction(2/4)

When using local support regions, it is implicitly assumed that all pixels in the region are of the same depth. the fronto-parallel surfaces assumption Adaptive-weight methods

5

Introduction(3/4)

Adaptive-weight methods are the local algorithms yielding the best results. Highly time-consuming task

Anisotropic diffusion, a computer vision technique very similar to adaptive weighting but computationally less expensive. a computer vision technique very similar to adaptive

weighting but computationally less expensive.

6

Introduction(4/4)

Geodesic diffusion is inspired by anisotropic diffusion. diffusing both matching costs and weights.

Near real-time execution is demonstrated using a commercial graphics card.

7

Related Work

Adaptive-weight methods [7] Adaptive-weight methods [8] Anisotropic diffusion [9]

[7] K.-J. Yoon and I.S. Kweon, “Adaptive Support-Weight Approach forCorrespondence Search,” IEEE Trans. Pattern Analysis and MachineIntelligence, vol. 28, no. 4, pp. 650-656, Apr. 2006.

[8] A. Hosni, M. Bleyer, M. Gelautz, and C. Rhemann, “Local Stereo MatchingUsing Geodesic Support Weights,” Proc. Int’l Conf. Image Processing,pp. 2093-2096, 2009.

[9] P. Perona and J. Malik, “Scale-Space and Edge Detection Using AnisotropicDiffusion,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 12,no. 7, pp. 629-639, July 1990.

8

Related Work(1/6)

Adaptive-weight methods [7]

9

Related Work(2/6)

Adaptive-weight methods [7]

123

123

123 truncated absolute difference (TAD)

Euclidean distance between the values in the CIELab color space and spatial euclidean distance

10

Related Work(3/6)

Adaptive-weight methods [8]

11

Related Work(4/6)

Adaptive-weight methods [8]

123

123

123

Shortest path

12

Related Work(5/6)

The two algorithms use the same optimization technique, winner-takes-all (WTA).

13

Related Work(6/6)

Anisotropic diffusion is a computer vision technique similar to bilateral filtering. only the comparison of each pixel with its immediate

neighbors is necessary.

14

Method

A : Anisotropic diffusion

B : Geodesic diffusion

15

Method.A(1/3)

Anisotropic diffusion

16

Method.A(2/3)

Anisotropic diffusion

123

123Euclidean distance between the values in the CIELab color space

17

Method.A(3/3)

It is an iterative computer vision technique.[9]

18

Method

A : Anisotropic diffusion

B : Geodesic diffusion

19

Method.B(1/8)

Three principles Costs and weights are diffused so that the importance of

each cost value is known in each iteration. In each iteration, the costs and weights at each pixel are

accumulated. After the last iteration, all the support region information has been accumulated at each pixel.

To increase the efficiency of information diffusion and to avoid loops, turns in the direction of diffusion are penalized.

20

Method.B(2/8)

Geodesic diffusion

21

Method.B(3/8)

Each of the four positions inherits the costs and weights of each of the four direct neighbors of each pixel.

22

Method.B(4/8)

Geodesic diffusion 123

123

123

23

Method.B(5/8)

i = 0 right neighbors i = 1 lower neighbors i = 2 upper neighbors i = 3 left neighbors

24

Method.B(6/8)

The cost and weight information derived from a direct neighbor is not returned to this neighbor.

Costs are only propagated with their full weights in the same direction of their propagation direction in the previous iteration.

25

Method.B(7/8)

Geodesic diffusion 123

123

26

Method.B(8/8)

At the end of the diffusion process, the DSI costs are normalized.

Thus, concluded, and the disparity map is then computed by selecting the lower cost disparity for each pixel WTA.

27

Experimental Results(1/8)

28

Experimental Results(2/8)

29

Experimental Results(3/8)

30

Experimental Results(4/8)

31

Experimental Results(5/)

32

Experimental Results(6/8)

33

Experimental Results(7/8)

34

Experimental Results(8/8)

35

CUDA

CUDA implementation of our algorithm ran in less than 60 milliseconds for the Tsukuba stereo pair on a GeForce 480 GTX card.

36

Q & A