✉ Corresponding author
- 🤗 [08/2026] Model checkpoints of the SOTA methods and text data are released on Hugging Face.
- 💻 [08/2026] Source code is now publicly available.
- 🎉 [07/2026] Our paper was accepted by ACM Multimedia 2026 (ACM MM 2026).
This repository provides the official implementation of ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection, accepted by ACM Multimedia 2026 (ACM MM 2026).
Infrared Small Target Detection (IRSTD) aims to accurately segment weak and tiny targets from complex infrared backgrounds. Existing pure-vision methods rely solely on pixel-level information and often struggle to distinguish small targets from background clutter. Meanwhile, existing vision-language methods typically describe targets and backgrounds using a single textual prompt, overlooking their inherent semantic asymmetry and introducing feature optimization conflicts.
To address these limitations, we propose ADGNet, an Asymmetric Dual-text Guided Network for infrared small target detection. Its main components include:
- Asymmetric Dual-text Prompt (ADP): employs an abstract, image-independent target prompt and a detailed, image-dependent background prompt to provide dedicated semantic guidance.
- Asymmetric Dual-Branch Interaction (ADBI): separately constructs a Target Localization branch and a Background Suppression branch to enhance weak targets and suppress complex clutter.
- Adaptive Feature Aggregation (AFA): dynamically integrates the features produced by the two branches to achieve accurate target segmentation.
We also construct the Asymmetric Image-Text Infrared (AITIR) dataset by providing asymmetric text annotations for three widely used infrared small target datasets: IRSTD-1K, NUDT-SIRST, and SIRST. Extensive experiments demonstrate that ADGNet achieves competitive performance against 21 state-of-the-art methods.
This repository provides:
- 💻 Training and inference code for ADGNet
- 🏆 Model checkpoints of the SOTA methods
- 📝 AITIR asymmetric text annotations
- 📊 Evaluation scripts
Figure 1. Overall architecture of ADGNet.
ADGNet
├── Figs/ # Overall framework of ADGNet
├── datasets/
│ ├── IRSTD-1K/
│ │ ├── images/ # Original infrared images
│ │ ├── masks/ # Ground-truth masks
│ │ ├── img_idx/ # Training and testing splits
│ │ └── text/ # Target and background prompts
│ ├── NUDT-SIRST/
│ └── SIRST/
├── model/
│ ├── ADGNet.py # Main ADGNet architecture
│ ├── TL_Branch.py # Target Localization branch
│ ├── BS_Branch.py # Background Suppression branch
│ └── DualStreamFusion.py # Dual-branch interaction and AFA
│
├── dataset.py # Dataset loader and augmentation
├── train.py # Training, testing, and inference
├── net.py # Network and loss wrapper
├── loss.py # SoftIoU loss
├── metrics.py # IoU, Pd, and Fa evaluation metrics
├── roc.py # ROC and AUC evaluation
├── params.py # FLOPs, parameters, and FPS analysis
├── utils.py # Data processing and training utilities
├── requirements.txt # Python dependencies
├── README.md
└── LICENSEgit clone https://github.com/iLearn-Lab/MM26-ADGNet.git
cd MM26-ADGNetconda create -n ADGNet python=3.10 -y
conda activate ADGNetADGNet uses the pretrained CLIP ViT-B/16 model. You can download the model weights from either Hugging Face or ModelScope.
sudo apt update
sudo apt install git-lfs
git lfs install
# Option A: Download from Hugging Face
git clone https://huggingface.co/openai/clip-vit-base-patch16
# Option B: Download from ModelScope
git clone https://www.modelscope.cn/openai-mirror/clip-vit-base-patch16.gitpip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121
pip install -r requirements.txtWe provide the trained checkpoints of ADGNet on three infrared small target detection datasets.
| Dataset | IoU (%) ↑ | Pd (%) ↑ | Fa (10⁻⁶) ↓ | Checkpoint |
|---|---|---|---|---|
| IRSTD-1K | 72.38 | 93.20 | 4.10 | Download |
| NUDT-SIRST | 95.53 | 99.47 | 2.64 | Download |
| SIRST | 83.08 | 100.00 | 4.97 | Download |
We release the asymmetric text annotations used to construct the AITIR dataset.
| Dataset | Text Annotations | Download |
|---|---|---|
| IRSTD-1K | Fixed Target Prompt + Detailed Background Prompt | Download |
| NUDT-SIRST | ||
| SIRST |
Train ADGNet on the selected dataset:
python train.py \
--trainset "IRSTD-1K" \
--testset "IRSTD-1K" \
--dataset_dir "./datasets" \
--epochs 600 \
--batchSize 16 \
--num_workers 8 \
--mode trainReplace
IRSTD-1KwithNUDT-SIRSTorSIRSTwhen training on another dataset.
Evaluate a trained checkpoint and generate prediction maps:
python train.py \
--trainset "IRSTD-1K" \
--testset "IRSTD-1K" \
--dataset_dir "./datasets" \
--mode test \
--ckpt "./SOTA_pth/ADGNet_mIoU_72.38_IRSTD-1K.pth.tar"The following figure presents qualitative comparisons between ADGNet and representative SOTA infrared small target detection methods on IRSTD-1K, NUDT-SIRST, and SIRST.
Figure 2. Qualitative comparisons of different methods on the IRSTD-1K, NUDT-SIRST, and SIRST datasets.
If you find this project useful in your research, please consider citing our paper:
Please also consider checking out and citing our other related work:
This project is released under the Apache License 2.0.
You may use, modify, and distribute the code in accordance with the terms of the license. Please retain the original license and attribution notices in redistributed or modified versions.