Double Layer Multi-Threshold Extension improving the generic object proposal generation
Xu Chen, Cuifang Zhang, Fei Yan · 2016
Generic object proposal generator plays an essential role in promoting the speed and accuracy of subsequent classification because of its ability of determining a general location of generic object in the image. Nevertheless many current object proposal generators somehow have the problems of time consuming or localization bias. In this paper, a new approach, Double Layer Multi-Threshold Extension (DLMTE) is proposed, which substantially is a bounding box refinement process to improve the performance of the object-locating method. Firstly, the object proposals produced by Binarized Normed Gradients (BING) method is taken as the initial object bounding boxes. And for validation of our method, the Edge Boxes is also regarded as an input. After box alignment procedure employed in the Multi-Thresholding Straddling Expansion (MTSE) method, according to superpixel boundary cue and global search, our DLMTE is utilized to refine the aligned object proposals to increase the location accuracy with simple computation. Experiments on PASCAL VOC 2007 dataset represent a better performance compared with M-Bing and M-EB improved by MTSE method.