Discriminative Learning with Scale Decomposition for Person Detection

摘要

Person detection,which can locate the person regions in the image,continues to be a hot research topic in both computer vision and signal processing communities.However,detecting person at small scale remains a challenging problem due to the lack of discriminative details in the typical image at small scale.In this paper,we propose a decomposition mapping method which contains two subnets:encoder subnet and decoder subnet.Encoder subnet can exploit decomposition transformation for person regions from big scale to small scale.Decoder subnet reverses the process of the encoder subnet.We add deconvolution network to the decoder subnet to make up for the lost information and a discriminative mapping has been restructured to transform the person regions from the small scale to the big scale.Therefore,person-regions and background-regions can then be separated according to their decomposition positions in the new scale space.The proposed approach is evaluated on two challenging person datasets:Caltech dataset and the KITTI dataset.Compared with SAF R-CNN,the miss rate has been optimized by 3.96%on Caltech person dataset and the mean average precision has been optimized by 1.76%on KITTI person dataset.