Abstract | ||
---|---|---|
Most state-of-the-art object detection networks need region proposals in their two-step framework. Popular region proposal networks can provide hundred proposals with acceptable accuracy. In this paper, we introduce a Multiple Filters Region Proposal Network (MFRPN) that can change its structure with dataset. We calculate the suitable sizes of filters and use multiple filters with appropriate reference boxes to make the regression of coordinates of proposals more accurate. To illustrate the proposed MFRPN, we adopt the framework of Faster R-CNN [1] and replace the RPN with the MFRPN. As a result, we get 0.98% improvement in mean AP on PASCAL VOC 2007 and 1.45% on PASCAL VOC 2012. |
Year | DOI | Venue |
---|---|---|
2017 | 10.1007/978-981-10-7305-2_6 | Communications in Computer and Information Science |
Keywords | DocType | Volume |
Object detection,Multiple filters,Reference box,Region proposal | Conference | 773 |
ISSN | Citations | PageRank |
1865-0929 | 0 | 0.34 |
References | Authors | |
0 | 5 |
Name | Order | Citations | PageRank |
---|---|---|---|
Dingqian Zhang | 1 | 0 | 0.34 |
Hui Zhang | 2 | 403 | 71.41 |
Wanling Zeng | 3 | 0 | 0.34 |
Zhongxing Han | 4 | 1 | 1.03 |
Xiaohui Hu | 5 | 17 | 8.10 |