SIRST-5K: Exploring Massive Negatives Synthesis with Self-supervised Learning for Robust Infrared Small Target Detection
- Introduction
- Dependencies and Installation
- Dataset
- Codes Demos
- Usage
- Quantative Results
- Citations
- Acknowledgement
- Following DNANet
- Python == 3.8
- pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
- pip install scikit-image
- pip install tqdm
- pip install matplotlib
- pip install tensorboard==2.14.0
- pip install opencv-python==4.8.0.76
Download the dataset download dir models[Baidu Drive][Google Drive]. Currently, the available dataset are:
SIRST-5K
: The dataset synthesized using negatives generation strategies (Fig 2).
Noise Sampling
# Run Noise_Sampling.py directly
python codes/Noise_Sampling/Noise_Sampling.py
Noise displacement
# Run add_noise.py directly
python codes/Mix_Rot/add_noise.py
Negative
# Run rot_patch.py directly
python codes/Mix_Rot/rot_patch.py
# Run rot_mask.py directly
python codes/Mix_Rot/rot_mask.py
Our negative augmentation strategies can produce large amounts of challenging image data. You can download the SIRST-5K directly for training.
python train.py --base_size 256 --crop_size 256 --epochs 1500 --dataset [dataset-name] --split_method 50_50 --deep_supervision True --train_batch_size 16 --test_batch_size 16 --mode TXT
python test.py --base_size 256 --crop_size 256 --st_model [trained model path] --model_dir [model_dir] --dataset [dataset-name] --split_method 50_50 --deep_supervision True --test_batch_size 1 --mode TXT
python visulization.py --base_size 256 --crop_size 256 --st_model [trained model path] --model_dir [model_dir] --dataset [dataset-name] --split_method 50_50 --deep_supervision True --test_batch_size 1 --mode TXT
Model | mIoU (x10(2)) | Pd (x10(2)) | Fa (x10(6)) | |
---|---|---|---|---|
Ours | 92.78 | 98.84 | 2.735 | [Weights] |
If you find this project useful for your research, please consider citing our paper. ๐
@ARTICLE{10496142,
author={Lu, Yahao and Lin, Yupei and Wu, Han and Xian, Xiaoyu and Shi, Yukai and Lin, Liang},
journal={IEEE Transactions on Geoscience and Remote Sensing},
title={SIRST-5K: Exploring Massive Negatives Synthesis with Self-supervised Learning for Robust Infrared Small Target Detection},
year={2024},
publisher={IEEE}
doi={10.1109/TGRS.2024.3387125}
}
This project is build based on DNANet. We thank the authors for sharing their code.