fast cost-volume filtering for visual correspondence and beyond

1

Fast Cost-volume Filtering For Visual Correspondence and Beyond

Asmaa Hosni, Member, IEEE,Christoph Rhemann,

Michael Bleyer, Member, IEEE,Carsten Rother, Member, IEEE,

and Margrit Gelautz,Senior Member, IEEE

IEEE Transactions on Pattern Analysis and Machine Intelligence,2013

2

Outline• Introduction• Related Work• Stereo Matching• Optical Flow• Interactive Image Segmentation

• Method ： Cost-Volume Filtering• Applications• Experimental Results• Conclusion

3

Introduction

4

Introduction• Many computer vision tasks can be formulated as labeling

problems.• Solution : a spatially smooth labeling where label transitions are

aligned with color edges. => very fast edge preserving filter• A generic and simple framework • (i) constructing a cost volume • (ii) fast cost volume filtering • (iii) winner-take-all label selection

5

Related Work

6

Related Work--Stereo Matching• Global method• Energy minimization process (GC,BP,DP,Cooperative)• Per-processing (mean-sift)• Accurate but slow

• Local method• A local support region with winner take all• Fast but inaccurate.

7

Related Work--Optical Flow• Optical flow ： the pattern of apparent motion caused by the

relative motion between an observer and the scene.• A vector field subject to Image Brightness Constancy Equation

(IBCE) • Application ： motion detection, object segmentation, motion

compensated encoding, and stereo disparity measurement......

8

Related Work--Optical Flow• Each flow vector is a label and subpixel accuracy further

increases the label space • Method ： continuous optimization strategies (coarse-to-fine)• SSD• Local convolution with oriented Gaussians[13]• Local convolution with bilateral filter[14]• Adaptive support weights[15,16]

• Trade off search space (quality) against speed.[13] D. Tschumperle` and R. Deriche, “Vector-Valued Image Regularization with PDE’s: A Common Framework for Different Applications,” CVPR, 2003.[14] J. Xiao, H. Cheng, H. Sawhney, C. Rao, and M. Isnardi, “Bilateral Filtering-Based Optical Flow Estimation with Occlusion Detection,” ECCV, 2006.[15] M. Werlberger, T. Pock, and H. Bischof, “Motion Estimation with Non-Local Total Variation Regularization,” CVPR, 2010.[16] D. Sun, S. Roth, and M. Black, “Secrets of Optical Flow Estimation and Their Principles,” CVPR, 2010.

9

Related Work--Interactive Image Segmentation• Interactive Image Segmentation is a binary labeling problem. • Goal ： Separate the image into foreground and background regions

given some hints by the user.• Method ：• SNAKE• Geodesic morphological

operator [8,21]

• Alpha matte [5,22]

[5] K. He, J. Sun, and X. Tang, “Guided Image Filtering,” ECCV, 2010.[8] A. Criminisi, T. Sharp, and C. Rother, “Geodesic Image and Video Editing,” ACM Graphics, 2010.[21] A. Criminisi, T. Sharp, and A. Blake, “GeoS: Geodesic Image Segmentation,”ECCV, 2008.[22] C. Rhemann, C. Rother, J. Wang, M. Gelautz, P. Kohli, and P. Rott, “A Perceptually Motivated Online Benchmark for Image Matting,” CVPR, 2009.

10

Cost-volume Filtering

11

Edge-preserving filtering• Edge-preserving filtering methods• Weighted Least Square [Lagendijk et al. 1988]• Anisotropic diffusion [Perona and Malik 1990]• Bilateral filter [Aurich and Weule 95], [Tomasi and Manduchi 98]• Digital TV (Total Variation) filter [Chan et al. 2001]

12

Bilateral filter [6]

[6] K. Yoon and S. Kweon, “Adaptive Support-Weight Approach for Correspondence Search,” IEEE Trans. Pattern Analysis and Machine Intelligence, Apr. 2006.

13

Bilateral filter• Advantages • Preserve edges in the smoothing process• Simple and intuitive• Non-iterative

• Disadvantages • Complexity O(Nr2)• Gradient distortion

Preserves edges,but not gradients

14

Guided filter [5]

[5] K. He, J. Sun, and X. Tang, “Guided Image Filtering,” Proc. European Conf. Computer Vision, 2010.

15

Guided filter

Bilateral filter does not have this linear model

16

Guided filter•

17

Bilateral filter V.S. Guided filter

18

Guided filter• Edge-preserving filtering• Non-iterative• O(1) time, fast and non-approximate• No gradient distortion

Advantages of bilateral filter

Disadvantages of bilateral filter

19

Cost-volume Filtering• Label l from

• • C’ : the filtered cost volume • i and j : pixel indices. • Wi,j : The filter weights depend on the guidance image I

•

• 　　　　： the mean vector and covariance of I• : a smoothness parameter• U : Identity matrix

• Winner take all ：k

r

r

w=(2r+1)^2

20


21[5] K. He, J. Sun, and X. Tang, “Guided Image Filtering,” Proc. European Conf. Computer Vision, 2010.[6] K. Yoon and S. Kweon, “Adaptive Support-Weight Approach for Correspondence Search,” IEEE Trans. Pattern Analysis and Machine Intelligence, Apr. 2006.[9] F. Crow. Summed-area tables for texture mapping. SIGGRAPH, 1984.


(a) Zoom of the green line in the input image. (b) Slice of the cost volume for the line(white/black/red: high/low/lowest costs)(c) The box filter [9](d) The joint bilateral filter [6](e) The guided filter [5](f) Ground-truth labeling

22

Application

23

Stereo Matching

24

Stereo Matching• Cost computation

• : grayscale gradients in x direction• : balances the color and gradient terms• : truncation values

25


• : grayscale gradients in x and y direction• : balances the color and gradient terms• : truncation values

• Occlusion detection :

26



• Occlusion detection :• Postprocessing

1.Scanline filling : the lowest disparity of the spatially closest nonoccluded pixel

2.Median filter :

• : adjust the spatial and color similarity• : normalization factor

27

Stereo Matching• Alternative—symmetric cost aggregation• The cost aggregation can be formulated symmetrically for both input

images. • Replace the 3*1 vector Ii in (3) with a 6*1 vector whose entries are given by

the RGB color channels of both Ii and I’i-l .

28

Stereo Matching

• Effect of postprocessing

Disparity map with invalidated pixels in red

After scanline-based filling After median filtering

29

Optical flow• Cost computation


• Postprocessing：Median filter and iteration

30

Optical flow• Upscale : To find subpixel accurate flow vectors.• Color of subpixel filled with bicubic interpolation.

31

Interactive Image Segmentation• 　　　　　： foreground F or background B• 　： background color histograms [25]• b(i) ： the bin of pixel i• Cost computation :

(binary labeling) =>• Five iterations [26]

[25] S. Vicente, V. Kolmogorov, and C. Rother, “Joint Optimization of Segmentation and Appearance Models,” Proc. IEEE Int’l Conf. Computer Vision, 2009.[26] C. Rother, V. Kolmogorov, and A. Blake, “Grabcut-Interactive Foreground Extraction Using Iterated Graph Cuts,” Proc. ACM Siggraph, 2004.

32

Experimental Results

33

Experimental Results• Device ： an Intel Core 2 Quad, 2.4 GHZ PC with GeForce

GTX480 graphics card with 1.5GB of memory from NVIDIA.• Settings parameters:

: depends on the signal-to-noise ratio of an image • 0.008 for stereo • 0.016 for optical flow

• Source : Middlebury http://vision.middlebury.edu/stereo/

http://vision.middlebury.edu/stereo/

34

Stereo Matching

35

Stereo Matching

36

Stereo Matching

• disparity maps without occlusion filling and postprocessing (invalidated pixels are black)

37

Stereo Matching

[27] A. Hosni, M. Bleyer, M. Gelautz, and C. Rhemann, “Local Stereo Matching Using Geodesic Support Weights,” Int’l Conf. Image Processing, 2009.[6] K. Yoon and S. Kweon, “Adaptive Support-Weight Approach for Correspondence Search,” PAMI , 2006.[7] C. Richardt, D. Orr, I. Davies, A. Criminisi, and N. Dodgson, “Real-Time Spatiotemporal Stereo Matching Using the Dual-Cross-Bilateral Grid,”ECCV,2010.

38

Stereo Matching• Million Disparity Estimations per second (MDE/s)

• W : width • H : height• D : disparity levels• FPS : number of frames per second

• A larger MDE number means a better performing system.

39

Stereo Matching

40

Stereo Matching

41

Stereo Matching

Table 2. Rankings and run times for selected local stereo methods. Run times in the table are averaged over the four Middlebury test images. *The run time in [15] was reported before left-right consistency check in the corresponding paper. Hence, for fairness, we have multiplied the reported time by a factor of 2.

42

Stereo Matching

43

Optical flow

44

Optical flow

The respective AEE and AAE are given in parentheses (AEE/AAE).

45

Optical flow

Comparison of two sequences with thin structure, where many competitors fail to preserve flow discontinuities. (We boosted the colors in the second row from the top for better visualization.)

46

Optical flowLarge displacement flow• (b)-(e) and (l)-(o) Motion magnitude for

different methods. • (f), (p) Our flow vectors with the color

coding as in middleburry. • (h)-(k) Backward warping results using

flow of different methods—the tip of the foot is correctly recovered by our method.

• Occluded regions cannot be handled by any method.

47

Interactive Image Segmentation

48


49


Our method GrabCut [26]Time 2.85 ms 300 mserror 5.3% 6.2%

[26] C. Rother, V. Kolmogorov, and A. Blake, “Grabcut-Interactive Foreground Extraction Using Iterated Graph Cuts,” Proc. ACM Siggraph, 2004.

50

Conclusion

51

Conclusion• Contribution• Propose a edge-preserving property • Runtime independent of the filter size • Achieved both accuracy and efficiency

• Future Work• Leverage this framework to other application areas • Process slanted surfaces [32]

[32] M. Bleyer, C. Rhemann, and C. Rother, “Patchmatch Stereo—Stereo Matching with Slanted Support Windows,” Proc. British Machine Vision Conf., 2011.

fast cost-volume filtering for visual correspondence and beyond

Documents