This is the official code for the paper "Tracker Meets Night: A Transformer Enhancer for UAV Tracking" accepted by IEEE RA-L with ICRA2022 presentation.

The spatial-channel Transformer (SCT) enhancer is a task-inspired low-light enhancer toward facilitating nighttime UAV tracking. Evaluations on the public UAVDark135 and the newly constructed DarkTrack2021 benchmarks demonstrate that the performance gains of SCT brought to nighttime UAV tracking surpass general low-light enhancers.

Paper | DarkTrack2021 benchmark

python 3.6
pytorch 1.8.1

Run lowlight_test.py, the results will be saved in ./result/

cd SCT
python lowlight_test.py 

Before training, you need to prepare the training set of the LOL dataset. Run lowlight_train.py. The model will be saved in ./log/SCT/models

cd SCT
python lowlight_train.py --trainset_path /your/path/to/LOLdataset/

To evaluate the performance of SCT in facilitating trackers' nighttime tracking ability, you need to meet the enviroment requirements of base trackers and download their snapshots to corresponding folders at first. Details can be found in their repos. Currently supporting trackers including HiFT, SiamAPN++, SiamRPN++, DiMP18, DiMP50, and PrDiMP50.

For HiFT, SiamAPN++, and SiamRPN++, change directory to their corresponding root, and simply run trackers with “--enhance” option

cd HiFT/SiamAPN++/pysot
python tools/test.py --dataset DarkTrack --enhance

For DiMP18, DiMP50, and PrDiMP50, customized your local paths in pytracking/evaluation/local.py

cd pytracking 
python run_tracker.py --tracker_name dimp --tracker_param dimp18/dimp50/prdimp50 --enhance 

The DarkTrack2021 benchmark comprises 110 challenging sequences with 100K frames in total. All sequences are captured at nighttime in urban scenes with a frame-rate of 30 frames/s (FPS). Some first frames of selected sequences in DarkTrack2021 are displayed below.

DarkTrack2021 is now available here.

If you find this work or code is helpful, please consider citing our paper:

@ARTICLE{Ye_2022_RAL,

author={Ye, Junjie and Fu, Changhong and Cao, Ziang and An, Shan and Zheng, Guangze and Li, Bowen},

journal={IEEE Robotics and Automation Letters},

title={{Tracker Meets Night: A Transformer Enhancer for UAV Tracking}},

year={2022},

volume={7},

number={2},

pages={3866-3873},

doi={10.1109/LRA.2022.3146911}

}

Junjie Ye Email: ye.jun.jie@tongji.edu.cn

Changhong Fu Email: changhongfu@tongji.edu.cn

A great thanks to Swin-Transformer for providing the basis for this code.