Pytorch implementation of our deep temporal probe network (DeepPro). [Paper]

Our work is currently under peer review. The contributions of this work are as follows:

• We reveal some new insights from a more crucial profile (i.e., temporal profile): long-term temporal information is much more essential for IRST detection, which includes global temporal saliency of target signals and correlation information between different signals. We validated the importance of the temporal profile in IRST detection, by developing the first predictive attribution tool.
• Inspired by our research, we remodel the IRST detection task as a one-dimensional signal anomaly detection task. Then, we propose an IRST detection network (i.e., DeepPro) to leverage the essential temporal profile information with multiply-add operations only in the time dimension.
• Experimental results show that DeepPro not only achieves a significant performance improvement in IRST detection with extremely high efficiency and continuity, but also has high robustness to dim targets and scenes with strong noise.

• Python 3
• torch
• tqdm
  
  

NUDT-MIRSDT   [download dir] is a synthesized dataset, which contains 120 sequences. We use 80 sequences for training and 20 sequences for test. We divide the test set into two subsets according to their SNR ((0, 3], (3, 10)).

In the test set, targets in 8 sequences are so weak (SNR lower than 3). It is very challenging to detect these targets. The test set includes Sequence[47, 56, 59, 76, 92, 101, 105, 119].

Other datasets include NUDT-MIRSDT-HiNo dataset [download dir], IRSDT dataset and RGBT-Tiny dataset.

Training set   [download dir] includes 1001 sequences. Validation set   [download dir] includes 202 sequences. Test set   [img download dir] [mask download dir] includes 200 sequences.

python train.py --model 'DeepPro' --seqlen 40 --dataset [dataset name] --datapath [dataset path]
python train.py --model 'DeepPro-Plus' --seqlen 40 --dataset [dataset name] --datapath [dataset path]

python test.py --seqlen 40 --datapath [dataset path] --dataset [dataset name] --logpath [log path] --log_dir [trained model path]
python test.py --seqlen 40 --datapath './datasets/SatVideoIRSDT' --dataset 'SatVideoIRSDT' --logpath './log/' --log_dir 'SatVideoIRSDT__2025-07-22_19-41__SoftLoUloss_DeepPro-Plus_DataL40'  # test for SatVideoIRSTD challenge

The comparison results of computational complexity and computational efficiency are as follows,

on NUDT-MIRSDT (SNR≤3)

on NUDT-MIRSDT (all)

on NUDT-MIRSDT-Noise

on SatVideoIRSDT

@article{li2025probing,
  title={Probing Deep into Temporal Profile Makes the Infrared Small Target Detector Much Better},
  author={Li, Ruojing and An, Wei and Ying, Xinyi and Wang, Yingqian and Dai, Yimian and Wang, Longguang and Li, Miao and Guo, Yulan and Liu, Li},
  journal={arXiv preprint arXiv:2506.12766},
  year={2025}
}
@article{li2023direction,
  title={Direction-coded temporal U-shape module for multiframe infrared small target detection},
  author={Li, Ruojing and An, Wei and Xiao, Chao and Li, Boyang and Wang, Yingqian and Li, Miao and Guo, Yulan},
  journal={IEEE Transactions on Neural Networks and Learning Systems},
  year={2023},
  publisher={IEEE}
}

@article{li2026satvideodataset,
  title={Infrared video satellite aerial moving target detection dataset and its evaluation},
  author={Li, Ruojing and Li, Zhaoxu and Chen, Nuo and Guo, Gaowei and Dou, Zechao and Long, Zhengxing and Luo, Yihang and Zeng, Yaoyuan and Sheng, Weidong and Li, Boyang and others},
  doi={10.11834/jig.250536},
  journal={Journal of Image and Graphics},
  pages={1--15},
  year={2026}
}
@article{li2025dataset,
  title={Infrared video satellite aerial moving target detection dataset},
  author={Li, Ruojing and Zeng, Yaoyuan and Sheng, Weidong and Li, Boyang and Li, Zhaoxu and Chen, Nuo and Guo, Gaowei and Dou, Zechao and Long, Zhengxing and Luo, Yihang and others},
  doi={10.57760/sciencedb.j00240.00077},
  url={https://doi.org/10.57760/sciencedb.j00240.00077},
  year={2025},
  publisher={Science Data Bank}
}

Welcome to raise issues or email to [email protected] for any question.