TheFastestDeformablePartModelforObjectDetection,-.PDFVIP

The Fastest Deformable Part Model for Object Detection Junjie Yan Zhen Lei Longyin Wen Stan Z. Li ? Center for Biometrics and Security Research National Laboratory of Pattern Recognition Institute of Automation, Chinese Academy of Sciences, China {jjyan,zlei,lywen,szli}@ Abstract in Pascal VOC. The speed is a bottleneck of DPM in real application, where speed is often as important as accuracy. This paper solves the speed bottleneck of deformable Recent works have accelerated deformable part model part model (DPM), while maintaining the accuracy in de- (DPM) by one order of magnitude, such as cascade [10], tection on challenging datasets. Three prohibitive steps in coarse-to-?ne [22], branch-and-bound [16] and FFT [8]. In cascade version of DPM are accelerated, including 2D cor- DPM, the detection score of each hypothesis is determined relation between root ?lter and feature map, cascade part by the score of appearance minus the deformation cost. The pruning and HOG feature extraction. For 2D correlation, appearance score is calculated by the correlation between the root ?lter is constrained to be low rank, so that 2D cor- HOG feature and a sequence of ?lters including root and relation can be calculated by more ef?cient linear combi- parts, which takes most of the time due to the high dimen- nation of 1D correlations. A proximal gradient algorithm is sion. [10] and [22] reduced the computation by pruning un- adopted to progressively learn the low rank ?lter in a dis- promising hypothesis early. [8] used FFT to accelerate cor- criminative manner. For cascade part pruning, neighbor- relation. These methods, however, still take about 1 second hood aware cascade is proposed to capture the dependence per image for Pascal VOC detection. We take cascade DPM