We are motivated by the need for a generic object proposal generationalgorithm which achieves good balance between object detection recall, proposallocalization quality and computational efficiency. We propose a novel objectproposal algorithm, BING++, which inherits the virtue of good computationalefficiency of BING but significantly improves its proposal localizationquality. At high level we formulate the problem of object proposal generationfrom a novel probabilistic perspective, based on which our BING++ manages toimprove the localization quality by employing edges and segments to estimateobject boundaries and update the proposals sequentially. We propose learningthe parameters efficiently by searching for approximate solutions in aquantized parameter space for complexity reduction. We demonstrate thegeneralization of BING++ with the same fixed parameters across different objectclasses and datasets. Empirically our BING++ can run at half speed of BING onCPU, but significantly improve the localization quality by 18.5% and 16.7% onboth VOC2007 and Microhsoft COCO datasets, respectively. Compared with otherstate-of-the-art approaches, BING++ can achieve comparable performance, but runsignificantly faster.
展开▼
