Presentation is loading. Please wait.

Presentation is loading. Please wait.

Unsupervised Learning of Categorical Segments in Image Collections *California Institute of Technology **Technion Marco Andreetto*, Lihi Zelnik-Manor**,

Similar presentations


Presentation on theme: "Unsupervised Learning of Categorical Segments in Image Collections *California Institute of Technology **Technion Marco Andreetto*, Lihi Zelnik-Manor**,"— Presentation transcript:

1 Unsupervised Learning of Categorical Segments in Image Collections *California Institute of Technology **Technion Marco Andreetto*, Lihi Zelnik-Manor**, Pietro Perona* The Sixth IEEE Computer Society Workshop on Perceptual Organization in Computer Vision (POCV 2008)

2 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

3 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

4 Motivation

5 Normalized cuts: Shi and Malik PAMI 2000

6 Motivation

7

8 Categorical segments: from human segmentation

9 Motivation

10 Related works Russell et al. CVPR 2006 Cao and Fei-Fei ICCV 2007 Wang and Grimson NIPS 2007 Andreetto et al. ICCV 2007

11 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

12 An image as a set of segments

13 K = 2 N

14 An image as a set of segments K = 2  Segment probability N

15 An image as a set of segments K = 2  Segment probability fkfk K Segment density N

16 Image formation K = 2  c Segment probability Label fkfk K x Segment density N

17 What we’re looking forObserved Probabilistic model for clustering  c fkfk KN x Likelihood of x to be in cluster k

18 Non-parametric densities Sum of local kernels

19 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

20 N Learning categorical segments  c fkfk K x w gkgk K Segment appearance Joint for all images Segment shape/color Specific per image M

21 Visual words Filter Bank VQ w1w1 w2w2 w3w3 wNwN Filter bank: 17 outputs 256 visual words Winn et al. ICCV 2005 … …

22 Inference N  c fkfk K x w gkgk K M

23 Gibbs sampling

24 Prior term: Number of pixels in image m assigned to segment k

25 Gibbs sampling Prior term: Visual words term: Number of pixels in image m assigned to segment k Number of visual word h assigned to segments k

26 Gibbs sampling Prior term: Visual words term: Segment term: Number of pixels in image m assigned to segment k Number of visual word h assigned to segments k Non-parametric density Estimate for segment k Affinity between observations i and j

27 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

28 Experimental results (MSRC)

29 Classification results (MSRC) Class NameWang and GrimsonOur model DetectionFalse Al.DetectionFalse Al. Cow0.56620.03340.48890.0823 GrassN/A 0.63890.0337 Cars0.68380.24370.33130.1732 SkyN/A 0.99540.0096 FoliageN/A 0.47350.1122 SeaN/A 0.61990.0174 Bikes0.56610.37140.54360.0646 Faces0.69730.42170.61610.0429 Running time: 18.75 sec. per image

30 Experimental results (Labelme)

31 Categorical segments (Labelme) Segment 1: FoliageSegment 2: Buildings Segment 1: SkySegment 3: Street pavement

32 Categorical segments (scenes)

33 Outline Motivation and related work A probabilistic model for single image segmentation Unsupervised learning of categorical segments Experimental results Conclusions and future works

34 Conclusions We presented a model for unsupervised learning of categorical segments We describe an inference method based on Gibbs sampling We show some experimental results on a standard dataset MSRC v1.

35 Future work Faster inference method (variational approximation) Automatic inference of the number of segments Learning geometric relationships between segments

36 Thank You

37


Download ppt "Unsupervised Learning of Categorical Segments in Image Collections *California Institute of Technology **Technion Marco Andreetto*, Lihi Zelnik-Manor**,"

Similar presentations


Ads by Google