基于图像联合分割与联合标注的服装解析问题研究.pptx

What makes for effective detection proposals?Reporter: Liliang ZhangJan et al. What makes for effective detection proposals? TPAMI15第页 | 共32页Object Proposal: IntuitionObject Proposal (OP) means the position ( or bbox) which may likely contain an object Prefer recall than precision since we can use a strong classifier to reject the false ones.第页 | 共32页OutlineOverview SelectiveSearch, CPMC, MCG Bing, Edgeboxes Multibox, DeepBoxDiscussion on RepeatabilityDiscussion on RecallDiscussion on Detection PerformanceConclusion第页 | 共32页OutlineOverviewDiscussion on RepeatabilityDiscussion on RecallDiscussion on Detection PerformanceConclusion第页 | 共32页Overview of different OP methodsGrouping proposal methods split the image into little pieces group and merge them to be the OPsWindow scoring proposal methods generate a set of bbox score the bbox by an (objectness) score rank the bbox to be the OPsAlternate proposal methodsOther methods, like the CNN based Multibox第页 | 共32页Grouping proposal methods第页 | 共32页SP: Grouping SuperpixelSelectiveSearchRandomizedPrimRantalankilaChangGC: Solving multiple graph cut problems with diverse seedsCPMCEndresRigorEC: Grouping directly from edge contoursGeodesicMCGSelective Search: Overview(1) Use [13] to get the smallest segment(2) Use multi-strategy grouping to merge the two segments with the max similarity to form a big segment(3) on and on until all segments merge to the whole image.(4) rank all the segments and generate the bboxes Jasper et al. Selective Search for Object Recognition. IJCV13[13] P. F. Felzenszwalb and D. P. Huttenlocher. Efficient GraphBased Image Segmentation. IJCV, 59:167–181, 2004. 1, 3, 4, 5, 7第页 | 共32页Constrained Parametric Min-Cuts(CPMC): OverviewStep1. General the fg and bg seed pixelsStep2. GraphCut under different degree of bias Step3. Rank the results according to a complex featureJ. Carreira and C. Sminchisescu, “Constrained parametric min-cuts forautomatic object segmentation,” in CVPR, 2010.第页 | 共32页Multiscale Comb