EDGSP

This is the code of paper 'Beyond Full Labels: Energy-Double-Guided Single-Point Prompt for Infrared Small Target Label Generation'

Shuai Yuan, Hanlin Qin, Renke Kou, XiangYan, Zechuan Li, Chenxu Peng, Huixin Zhou [Paper] [Weight]

Chanlleges and inspiration

Structure

Introduction

We present a novel infrared small target label generation (IRSTLG) framework named energy double guided single-point prompt (EDGSP). Experiments on both public (e.g., SIRST, NUDT-SIRST, IRSTD-1K) demonstrate the effectiveness of our method. Our main contributions are as follows:

To the best of our knowledge, we present the first study of the learning-based IRSTLG paradigm and introduce EDGSP creating a crucial link between label generation and target detection task.
We propose target energy initialization (TEI), double prompt embedding (DPE), and bounding box-based matching (BBM) strategies to address insufficient shape evolution, label adhesion, and false alarms.
For the first time, three baselines equipped with EDGSP achieve accurate annotation on three datasets. The downstream detection task illustrates that our pseudo label surpasses the full label. Even with coarse point annotated, EDGSP achieves 99.5% performance of full labeling.

If the implementation of this repo is helpful to you, just star it！⭐⭐⭐

Usage

1. Data

Note that using the “fixed” file to correct seven obvious errors in the raw data！！！
SIRST3: SIRST, NUDT-SIRST, and IRSTD-1K

Our project has the following structure:

├──./datasets/
│    ├── SIRST3
│    │    ├── images
│    │    │    ├── XDU0.png
│    │    │    ├── XDU1.png
│    │    │    ├── ...
│    │    ├── masks
│    │    │    ├── XDU0.png
│    │    │    ├── XDU1.png
│    │    │    ├── ...
│    │    ├── Centroid
│    │    │    ├── XDU0.png
│    │    │    ├── XDU1.png
│    │    │    ├── ...
│    │    ├── masks_coarse
│    │    │    ├── XDU0.png
│    │    │    ├── XDU1.png
│    │    │    ├── ...
│    │    ├── img_idx
│    │    │    ├── train_SIRST3.txt
│    │    │    ├── test_SIRST3.txt

2. Train.

python train_LG_SCTrans.py

3. Test and demo.

python test_LG_SCTrans_PdFa.py

Results and Trained Models

Qualitative Results

Quantitative Results on Mixed Dataset (SIRST3): SIRST, NUDT-SIRST, and IRSTD-1K

Model	mIoU (x10(-2))	Pd (x10(-2))	Fat (x10(-2))	Fa (x10(-6))
SIRST	83.83	100.0	0	0
NUDT-SIRST	95.51	100.0	0	0
IRSTD-1K	73.80	100.0	0	0
[Weights]

*This code is highly borrowed from IRSTD-Toolbox. Thanks to Xinyi Ying.

*The overall repository style is highly borrowed from DNA-Net. Thanks to Boyang Li.

Contact

Welcome to raise issues or email to yuansy@stu.xidian.edu.cn or yuansy2@student.unimelb.edu.au for any question regarding our EDGSP.

xdFai/EDGSP