iLearn-Lab/MM26-ADGNet

ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection

Tongtong Wang¹  Mingzhu Xu^1✉  Chenglong Yu¹  Jing Wang¹  Xiaohui Lin¹  Weili Guan²

¹School of Software, Shandong University
²Harbin Institute of Technology, Shenzhen
^✉Corresponding author

📖 Model Description

This repository provides the official model checkpoints for ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection, accepted by ACM Multimedia 2026.

Infrared Small Target Detection aims to accurately segment weak and tiny targets from complex infrared backgrounds. Existing pure-vision methods rely mainly on pixel-level information, while existing vision-language methods commonly describe targets and backgrounds using a single textual prompt. Such a symmetric design overlooks the inherent semantic differences between sparse infrared targets and structurally complex backgrounds.

ADGNet addresses this problem through three main components:

• Asymmetric Dual-text Prompt (ADP): uses an abstract, image-independent target prompt and a detailed, image-dependent background prompt.
• Asymmetric Dual-Branch Interaction (ADBI): independently performs target localization and background suppression using their corresponding textual priors.
• Adaptive Feature Aggregation (AFA): dynamically fuses target-enhanced and background-suppressed features for accurate segmentation.

The model uses the pretrained CLIP ViT-B/16 text encoder to extract semantic representations from the target and background prompts.

---

🏆 Available Checkpoints

All ADGNet checkpoints are hosted in this Hugging Face model repository.

Download the required checkpoint directly from the Files and versions section of this repository.

Dataset	Checkpoint
IRSTD-1K	ADGNet_mIoU_72.38_IRSTD-1K.pth.tar
NUDT-SIRST	ADGNet_mIoU_95.53_NUDT-SIRST.pth.tar
SIRST	ADGNet_mIoU_83.08_SIRST.pth.tar

---

🚀 Usage

These checkpoints are designed to be used with the official ADGNet implementation:

https://github.com/iLearn-Lab/MM26-ADGNet

1. Clone the Official Repository

git clone https://github.com/iLearn-Lab/MM26-ADGNet.git
cd MM26-ADGNet

2. Prepare the Checkpoints

Place the downloaded checkpoints in:

MM26-ADGNet/
└── SOTA_pth/
    ├── ADGNet_mIoU_72.38_IRSTD-1K.pth.tar
    ├── ADGNet_mIoU_95.53_NUDT-SIRST.pth.tar
    └── ADGNet_mIoU_83.08_SIRST.pth.tar

3. Prepare the CLIP Text Encoder

ADGNet uses the pretrained CLIP ViT-B/16 model:

git clone https://huggingface.co/openai/clip-vit-base-patch16

Update the local CLIP model path in the corresponding project configuration or source file before inference.

4. Run Evaluation

Example evaluation on IRSTD-1K:

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"

Replace the dataset name and checkpoint path when evaluating on NUDT-SIRST or SIRST.

---

📂 Dataset and Text Annotation Preparation

The original infrared images and ground-truth masks are not included in this model repository. Please obtain IRSTD-1K, NUDT-SIRST, and SIRST from their respective official sources.

The asymmetric text annotations used by ADGNet are released separately in our Hugging Face dataset repository:

• AITIR Text Annotations: Download

After downloading the original datasets and text annotations, organize them according to the official ADGNet repository:

datasets/
├── IRSTD-1K/
│   ├── images/
│   ├── masks/
│   ├── img_idx/
│   └── text/
├── NUDT-SIRST/
│   ├── images/
│   ├── masks/
│   ├── img_idx/
│   └── text/
└── SIRST/
    ├── images/
    ├── masks/
    ├── img_idx/
    └── text/

---

🎯 Intended Use

The released checkpoints are intended for:

• Academic research on infrared small target detection
• Reproduction of the results reported in the ADGNet paper
• Evaluation on IRSTD-1K, NUDT-SIRST, and SIRST
• Research on multimodal and text-guided infrared image segmentation
• Comparison with other infrared small target detection methods

---

⚠️ Limitations

• The model requires both infrared images and corresponding textual prompts.
• Detection performance may vary when applied to datasets or scenes that differ substantially from the training distribution.
• The released checkpoints are designed for the dataset splits and evaluation settings used in the paper.
• The model depends on the pretrained CLIP ViT-B/16 text encoder.
• The original infrared datasets are subject to their respective licenses and terms of use.

---

🔗 Related Resources

• GitHub Repository: iLearn-Lab/MM26-ADGNet
• Paper: ADGNet: Asymmetric Dual-text Guided Network for Infrared Small Target Detection
• Text Annotations: AITIR Text Annotations

---

📚 Citation

If you find ADGNet or the released checkpoints useful in your research, please consider citing our paper:

Please also consider checking out and citing our other related work:

---