Abstract | ||
---|---|---|
The multi-task learning framework that considers pedestrian detection and person re-identification jointly is an effective solution for person search. However, the existing joint frameworks simply share the backbone network without considering the negative interaction between the two tasks. To alleviate this conflict and meet the different requirements in detection and re-identification, a Partially Separated Network (PSN) for person search is proposed in this paper. Unlike the traditional joint frameworks, our backbone network is partially separated for detection and identification, and feature maps with different scales are provided according to different characteristics. Our experiment results have demonstrated that on CUHK-SYSU dataset our mAP and top-1 on ResNet-50 are 5.4% and 4.4% higher, and on PRW dataset our mAP and top-1 on PVANet are 8.0% and 5.0% higher compared with the state-of-the-art methods. Specially, the improvements can be more impressive in the case of large gallery, occlusion and low resolution. |
Year | DOI | Venue |
---|---|---|
2018 | 10.1007/978-3-030-00764-5_71 | ADVANCES IN MULTIMEDIA INFORMATION PROCESSING, PT III |
Keywords | Field | DocType |
Person search,Pedestrian detection,Person re-identification,Multi-task learning,Partially separated network | Computer vision,Multi-task learning,Pattern recognition,Computer science,Artificial intelligence,Backbone network,Pedestrian detection | Conference |
Volume | ISSN | Citations |
11166 | 0302-9743 | 0 |
PageRank | References | Authors |
0.34 | 20 | 4 |
Name | Order | Citations | PageRank |
---|---|---|---|
Chuanchuan Chen | 1 | 0 | 0.68 |
Jingbo Fan | 2 | 0 | 0.68 |
Zhu Yuesheng | 3 | 112 | 39.21 |
Luo Guibo | 4 | 15 | 6.04 |