This paper has proposed a tracker for single object tracking from view of UAVs. Excellent performance has been validated across 5 commonly adopted benchmarks, i.e UAV123_10fps, UAV20L, DTB70, VisDrone2020 and LaTOT_test. The training datasets contain ImageNet DET, ImageNet VID, COCO, LaSOT, GOT-10K.
Three main contributions are attributed for TSN's outstanding tracking performance.
- Size-Aware Module (SAM): Leverage features extensively through multi-receptive learning to adopt size variations of tracked object.
- Center Attention Module (CAM): Suppress background distractions and highlight object region for consistent tracking.
- Dynamic Positive sample Definition Strategy (DPDS): Adjust positive sampling area subject to scale variation in the training stage for better convergence.
The implementation is based on PySOT and SiamCAR. Questions about environment creation, testing and et. al, can be firstly referred to their official repositories and blogs, which we have found them helpful enough. Corresponding result files of TSN on different benchmarks have been uploaded. Codes of network definition, benchmark evaluation, related checkpoints and hyper-params are available at present.
| Result text files | Snapshots for TSN |
|---|---|
| 百度网盘 | 百度网盘 |
Some visualization of TSN along with Ground Truth annotations are provided. Since UAV20L and VisDrone2019 are too long to be visulaized due to their long-term tracking scenes, we only display tracking results of relatively-short videos from the other 3 datasets.
Animal2.mp4
SpeedCar2.mp4
car16.mp4
car26.mp4
person14_3.mp4
wakeboard7.mp4
If you find this repository helpful, please consider citing this paper:
@article{liang2024target,
title={Target signature network for small object tracking},
author={Liang, Lei and Chen, Zhihua and Dai, Lei and Wang, Shouli},
journal={Engineering Applications of Artificial Intelligence},
volume={138},
pages={109445},
year={2024},
publisher={Elsevier}
}