Improved Global-sampling Matting Using Sequential Pair-selection Strategy

Transcription

1 IS&T/SPIE Electronic Imaging 2014 University of Ottawa School of Electrical Engineering and Computer Science Improved Global-sampling Matting Using Sequential Pair-selection Strategy Presented By: Ahmad Al-Kabbany Under the Supervision of: Prof.Eric Dubois February 6, 2014

2 Outline Sampling-based alpha matting State-of-the-art sampling-based matting Sequential pair-selection matting Future directions 2

3 Problem statement Hard Segmentation vs. Soft Foreground/Background Segmentation Source: A. Levin, A. Rav-Acha, and D. Lischinski. Spectral matting.ieee TPAMI, 30: , October 2008 Propagation-based matting Sampling-based matting 3

4 The linear convex image model (compositing equation) (1) (2) (3) 4

5 The linear convex image model (compositing equation) (1) (2) (3) The trimap 5

6 The Trimap Source: A. Levin, A. Rav-Acha, and D. Lischinski. Spectral matting.ieee TPAMI, 30: , October 2008 Manually generated Heavily affects the output of all existing algorithms 6

7 Sampling-based Matting Sampling a search pool of FG/BG pairs. Usually searching nearby regions in the trimap Picking the best pair Sampling strategies Pair-selection objective functions The distant-but-true problem 7

8 The distant-but-true problem June 18,

9 Global Sampling Matting Reference: K.He, C.Rhemann, C.Rother, X.Tang and J.Sun. A Global Sampling Method for Alpha Matting, CVPR,

10 Comprehensive Sampling Matting Source: E. Shahrian, D. Rajan, B. Price and S. Cohen. Improving Matting Using Comprehensive Sampling Sets, CVPR,

11 Sequential Pair-selection (SPS) Matting 11

12 A few drawbacks of existing techniques Color ambiguity problem 12

13 A few drawbacks of existing techniques Color ambiguity problem Distance as a criterion to augment search pool An intuitive observation: There is something true nearby, otherwise it is an isolated region Picking a true half-pair 13

14 Over-segmentation Ray shooting 14

15 Over-segmentation Ray shooting RB2 RB1 RFm Rl 15

16 Over-segmentation Ray shooting Cohen's d-value checks 16

17 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling 17

18 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling 1. First leaf: S samples will be taken from particular near FG and BG regions 18

19 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling 1. First leaf: S samples will be taken from particular near FG and BG regions 2. Second leaf: FG samples will be sought nearby while knn non-local BG samples will be sought from all known BG regions 19

20 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling 1. First leaf: S samples will be taken from particular near FG and BG regions 2. Second leaf: FG samples will be sought nearby while knn non-local BG samples will be sought from all known BG regions 3. Third leaf: BG samples will be sought nearby while knn non-local FG samples will be sought from all known FG regions 20

21 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling 1. First leaf: S samples will be taken from particular near FG and BG regions 2. Second leaf: FG samples will be sought nearby while knn non-local BG samples will be sought from all known BG regions 3. Third leaf: BG samples will be sought nearby while knn non-local FG samples will be sought from all known FG regions 4. Fourth leaf: occasionally happens if FG and BG color distributions overlap 21

22 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling Pair space streaking 22

23 Over-segmentation Ray shooting Cohen's d-value checks knn non-local sampling Pair space streaking 23

24 24

25 Matting online benchmark (alphamatting.com) Three new records over all the matting techniques (particular image, trimap and metric) Records of the earlier variant of this work A wider variety of matting challenges Global-sampling matting (filter version) SAD MSE Gradient Connectivity 25

26 26

27 Advantages No color ambiguity (space streaking) The distance is not the measure for augmenting the search pool The search pool contains the best pair Drawbacks The search pool is relatively large The performance is poor when the FG and the BG color distributions overlap 27

28 Thank you 28