Infrared Small Target Detection in Satellite Videos: A New Dataset and A Novel Recurrent Feature Refinement Framework

IRSatVideo-LEO Dataset

IRSatVideo-LEO is a large-scale dataset for multi-frame infrared small target (MIRST) detection in LEO satellite videos. IRSatVideo-LEO is a semi-simulated dataset with a real satellite background image and synthesized satellite motion, target appearance, trajectory and intensity. IRSatVideo-LEO consists of 200 image sequences, 91021 frames and 218038 targets, and we provide instance-level segmentation annotations to offer a infrared LEO satellite videos benchmark for MIRST detection and tracking.

Downloads

[Baidu Yun DownLoads]

Implementation

Fig. 1 Implementation details of the IRSatVideo-LEO dataset.

Table 1 Details of parameters for data generation.

Benchmark Properties

Table 2 Statical comparisons among existing SIRST and MIRST detection datasets and our IRSatVideo-LEO dataset.

Fig. 2 Illustrations of sequence attributes.

Fig. 3 Illustrations of target attributes.

Fig. 4 Illustration of background attributes and example images.

Recurrent Feature Refinement Framework

Overview

Fig. 5 The proposed architecture of recurrent feature refinement framework.

Requirements

• Python 3.8
• pytorch (1.10.1+cu11.1), torchvision (0.11.2+cu11.1)

Build

DCN Compiling

1. Cd to ./codes/model/dcn.
2. Run bash make.sh. The scripts will build D3D automatically and create some folders.
3. See test.py for example usage.

Commands for Training

• Run train.py to perform network training. Example for training [model_name] on [dataset_name] datasets:

  $ cd ./codes
  $ python train.py --model_name ['ISTUDNet_RFR', 'ResUNet_RFR'] --dataset_name ['IRSatVideo-LEO']
  
• Checkpoints and Logs will be saved to ./codes/log/, and the ./codes/log/ has the following structure:

  ├──./codes/log/
  │    ├── [dataset_name]
  │    │    ├── [model_name]_20.pth.tar
  

Datasets

Download the IRSatVideo-LEO Dataset dataset and put the images in ./codes/data/IRSatVideo-LEO.

  IRSatVideo-LEO
    └── images
        ├── AfricaWest-1_38
            ├── 0000.png
            ├── 0001.png
            ├── ...
        ├── EastAfrica-0
            ├── 0000.png
            ├── 0001.png
            ├── ...
        ├── ...
    └── masks
        ├── AfricaWest-1_38
            ├── 0000.png
            ├── 0001.png
            ├── ...
        ├── EastAfrica-0
            ├── 0000.png
            ├── 0001.png
            ├── ...
        ├── ...		
    └── video_idx
        ├── train_IRSatVideo-LEO.txt
        ├── test_IRSatVideo-LEO.txt
        ├── test_IRSatVideo-LEO-easy.txt
        ├── test_IRSatVideo-LEO-middle.txt
        ├── test_IRSatVideo-LEO-hard.txt
    └── img_idx
        ├── AfricaWest-1_38.txt
        ├── EastAfrica-0.txt
        ├── ...      

Commands for Test

• Run test.py to perform network inference. Example for test [model_name] on [dataset_name] datasets:

  $ cd ./codes
  $ python test.py --model_name ['ISTUDNet_RFR', 'ResUNet_RFR'] --dataset_name ['IRSatVideo-LEO'] --save_img True
  

• The PA/mIoU and PD/FA values of each dataset will be saved to ./test_[current time].txt
  

• Network preditions will be saved to ./results/ that has the following structure:

  ├──./results/
  │    ├── [dataset_name]
  │    │   ├── [model_name]
  │    │   │    ├── 0000.png
  │    │   │    ├── 0001.png
  │    │   │    ├── ...
  

Results

Table 3 $P_d$ ($\times {10}^{-2}$), $F_a$($\times {10}^{-6}$) values achieved by different methods on IRSatVideo-LEO dataset.

Fig. 6 Qualitative results of different methods.

Contact

Please contact us at [email protected] for any questions.

Citiation

@article{RFR,
  author = {Xinyi Ying, Li Liu, Zaipin Lin, Yangsi Shi, Yingqian Wang, Ruojing Li, Xu Cao, Boyang Li, Shilin Zhou},
  title = {Infrared Small Target Detection in Satellite Videos: A New Dataset and A Novel Recurrent Feature Refinement Framework},
  journal = {Arxiv},
  year = {2024},
}