Region-based Voting
Exemplar 1
Query
1
Exemplar 2
Region-based Voting
Query
2
Exemplar 1
Exemplar 2
Region-based Voting
Query Query
MeanShift
Clustering
3
Exemplar 1
Exemplar 2
Computer Vision GroupUC Berkeley
Discriminative Weight Learning
• Not all regions are equally important
Frome, Singer and Malik. NIPS ‘06
image J exemplar I image K
want:
DIJ DIK
DIK > DIJMax-margin formulation results in a sparse solution of weights.
DIJ = Σi wi · diJand di
J=minj χ2(fiI, fj
J)
Computer Vision GroupUC Berkeley
Weight Learning Results
Algorithm Pipeline
Region matching based voting
Verification classifier
Constrained segmenter
Query
Exemplars
Images
Ground truths
Initial Hypotheses Segmentation
Detection
Weight learning
6
Initial Object/Background Labels
Initial Labels
Exemplar
7
Transformed Mask
Query Matched Part
: Object label: Background label: Unknown label
+
Fully automatic unlike interactive use of Graph Cuts, e.g. Blake et al. ECCV 04
Propagate Object/Background Labels
8
Arbelaez and Cohen. CVPR 08Initial Labels Final Segmentation
Computer Vision GroupUC Berkeley
ETHZ Shape (Ferrari et al. 06)• Contains 255 images of 5 diverse shape-based
classes.
Computer Vision GroupUC Berkeley
Detection Results on ETHZ
Hough baseline1 kAS 1 Shape 2 Ours
Det. rate at 0.3FPPI 31.0% 62.4% 67.2% 87.1±2.8%
1. Ferrari et al. PAMI 2008. 2. Ferrari, Jurie, Schmid. CVPR 2007
Computer Vision GroupUC Berkeley
Detection Results on ETHZ
Computer Vision GroupUC Berkeley
Detection Results on ETHZ
Computer Vision GroupUC Berkeley
Segmentation Results on ETHZ
Orig. Image Segmentation
The mean average precision is 75.7±3.2%
Orig. Image Segmentation
Computer Vision GroupUC Berkeley
Segmentation Results on ETHZ
Orig. Image Segmentation Orig. Image Segmentation
Computer Vision GroupUC Berkeley
Segmentation Results on ETHZ
Orig. Image Segmentation Orig. Image Segmentation
Computer Vision GroupUC Berkeley
Segmentation Results on ETHZ
Orig. Image Segmentation Orig. Image Segmentation
Computer Vision GroupUC Berkeley
Segmentation Results on ETHZ
Orig. Image Segmentation Orig. Image Segmentation
Computer Vision GroupUC Berkeley
Complexity Reduction
Computer Vision GroupUC Berkeley
Caltech 101 results
Computer Vision GroupUC Berkeley
Context from region tree (ICCV 09)
Computer Vision GroupUC Berkeley
MSRC dataset
Computer Vision GroupUC Berkeley
Confusion matrix (mean diagonal 67%)
Computer Vision GroupUC Berkeley
Concluding Remarks
• Our approach– Bottom up region segmentation– Hough transform style voting (learned weights)– Top down segmentation– Capture context by region tree
• Results on ETHZ , Caltech 101, MSRC competitive
• Lot more needs to be done to produce a robust solution to the problem of combining top down and bottom up information, but I think this is the central problem of vision