Here is the source code of our paper Ground-to-Aerial Person Search: Benchmark Dataset and Approach.
In this paper we are the first to construct a large-scale ground-to-aerial person search benchmark dataset, named G2APS, for the
cross-platform ground-to-aerial person search. The dataset can be downloaded at here code 1357.
We are glad to see research on G2APS and new ideas for Ground-to-Aerial Person Search.
Fig.1 Exemplars of aerial images and their corresponding ground surveillance images.
Besides, Head Knowledge Distillation(HKD) module is proposed in this paper to alliviate conflict exists in Optimization objective in detection and Re-ID subtasks.Finally, our method gain sota-of-the-art performance in G2APS,CUHK-SYSY and PRW.
Fig.2 Overview of our head knowledge distillation(HKD) framework.
Tab.1 Performance of baseline+HKD comparision with SOTA method.
| dataset | method | mAP | top-1 | model |
|---|---|---|---|---|
| G2APS | SeqNet | 33.96 | 44.52 | - |
| SeqNet+HKD | 39.40(+5.44) | 49.12(+4.60) | model | |
| COAT | 40.32 | 50.53 | - | |
| COAT+HKD | 41.41(+1.09) | 51.94(+1.41) | model | |
| CUHK-SYSU | SeqNet | 93.8 | 94.6 | - |
| SeqNet+HKD | 95.25(+1.45) | 96.10(+1.5) | model | |
| COAT* | 93.68 | 94.1 | - | |
| COAT+HKD | 93.86(+0.18) | 94.76(+0.66) | model | |
| PRW | SeqNet | 46.7 | 83.4 | - |
| SeqNet+HKD | 51.49(+4.79) | 85.12(+1.72) | model | |
| COAT* | 52.45 | 86 | - | |
| COAT+HKD | 53.49(+1.04) | 86.63(+0.63) | model |
note:* indicates the experimental results we reproduced.
Fig.3 Evaluation of two test samples.
If you find the code useful, please consider citing our paper:
@inproceedings{zhang2023ground,
title={Ground-to-Aerial Person Search: Benchmark Dataset and Approach},
author={Zhang, Shizhou and Yang, Qingchun and Cheng, De and Xing, Yinghui and Liang, Guoqiang and Wang, Peng and Zhang, Yanning},
booktitle={Proceedings of the 31st ACM International Conference on Multimedia},
pages={789--799},
year={2023}
}