Name
Abstract | ||
|---|---|---|
The background differences between domains poses a huge challenge for cross-domain person re-identification (Re-ID). And the attention mechanism is usually used to select regions of interest and suppress irrelevant background noises. However, many realizations of the attention mechanism do not deeply dig the relevance of the internal information of the features, and the attention information at different stages in the model is independent of each other. In response to the above problems, we propose the Spatial Attention Pyramid Network (SAPN), which can fuse the saliency information at different stages, thereby enhancing the model's adaptability to cross-domain Re-ID. First, the Instance Normalization (IN) layer is inserted in the backbone to eliminate the style differences of images by normalizing the data distribution. Secondly, a novel Spatial Attention Block (SAB) is designed to accurately locate the salient features of pedestrains and suppress background noises. Finally, we draw on the idea of Feature Pyramid Network (FPN) to design the Attention Embedded Pyramid Module (AEPM), which combines high-level semantic information with low-level location information, improves the cross-domain generalization capability of the model. We validate our SAPN through extensive experiments on three datasets which usually used in person Re-ID, the experiments show consistent improvements in a variety of cross-domain scenarios. |
Year | DOI | Venue |
|---|---|---|
2021 | 10.1007/978-3-030-86130-8_5 | WIRELESS ALGORITHMS, SYSTEMS, AND APPLICATIONS, WASA 2021, PT II |
Keywords | DocType | Volume |
Attention, Salient features, Background noises, Cross-domain person re-identification | Conference | 12938 |
ISSN | Citations | PageRank |
0302-9743 | 0 | 0.34 |
References | Authors | |
0 | 5 | |
Name | Order | Citations | PageRank |
|---|---|---|---|
| Zhaoqian Jia | 1 | 0 | 0.34 |
| Wenchao Wang | 2 | 0 | 0.34 |
| Shaoqi Hou | 3 | 0 | 1.01 |
| Ye Li | 4 | 61 | 7.26 |
| Guangqiang Yin | 5 | 2 | 5.79 |