🪄 AuxDet: Auxiliary Metadata Matters for Omni-Domain Infrared Small Target Detection

---

Authors: Yangting Shi, Renjie He, Le Hui, Xiang Li, Jian Yang, Ming-Ming Cheng, Yimian Dai

The official implementation of "AuxDet: Auxiliary Metadata Matters for Omni-Domain Infrared Small Target Detection".

---

🧾 Introduction

Infrared imaging plays a vital role in real-world applications such as marine resource exploration, precision guidance, and ecological monitoring, thanks to its ability to detect targets under harsh weather and low-visibility conditions.

However, most existing work in infrared small target detection (IRSTD) focuses narrowly on domain-specific feature engineering, while ignoring the wide-area, cross-domain detection needs of real-world infrared systems.

This has led to a serious issue: current IRSTD methods often break down when evaluated under Omni-Domain settings, where imaging conditions—such as platforms, spectral bands, and resolutions—vary significantly.

🎯 Our core research question is: In Omni-Domain IRSTD, can a model recognize the imaging domain of each input and adapt accordingly? Our answer is yes.

📡 In practical infrared sensing systems, each captured image comes with a set of auxiliary metadata—such as sensor platform, spectral band, and spatial resolution.

💡 While often overlooked, this metadata encodes key factors driving the domain-specific appearance and feature distribution. It can serve as a compact, semantically rich descriptor of the imaging domain.

---

🔍 Architecture

We propose the Auxiliary Metadata Driven Detector (AuxDet), a novel framework for Omni-IRSTD that redefines the conventional visual-only paradigm by leveraging auxiliary metadata for multi-modal scene-aware optimization.

---

📊 Dataset

We adopt the WideIRSTD-Full dataset as our evaluation benchmark. The dataset consists of 7 public datasets and an additional dataset developed by the team at the National University of Defense Technology (NUDT). It is important to note that auxiliary metadata corresponding to images is only provided in the training set. Therefore, for the training set only, we used the re-splitting provided in https://github.com/YeRen123455/ICPR-Track2-LightWeight/tree/main/dataset/ICPR_Track2/70_20 for our experiments.

---

🚀 Installation

Step 1. Clone the repository

git clone https://github.com/GrokCV/AuxDet.git
cd AuxDet

Step 2. Create environment and install dependencies

conda create --name AUXD python=3.9 -y
conda activate AUXD
conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia
pip install -U openmim
mim install mmengine  
mim install "mmcv>=2.0.0rc4,<2.2.0"
# cd AuxDet/
pip install -v -e .

---

🧊 Getting Started

First, place your dataset files according to the following folder structure.

data_root
    |--VOC2007
        |-- ImageSets
            |-- Main # Following the dataset split from https://github.com/YeRen123455/ICPR-Track2-LightWeight/tree/main/dataset/ICPR_Track2/70_20
                |-- train.txt
                |-- val.txt
        |-- Annotations
            |-- 00001.xml 
            |-- 00002.xml
            |-- ...
        |-- PNGImages # Images under WideIRSTD-Full Dataset/train/images (a total of 9,000 images)
            |-- 00001.png 
            |-- 00002.png
            |-- ... 

We convert the image bounding boxes and related auxiliary information into the XML format. You can download the annotations here.

Remember to update the dataset paths in the code accordingly.

# Open the file: ./configs/_base_/datasets/voc_aux.py

# Update Line 2 with your dataset root directory
data_root = "your own root directory" 

# Modify Line 60 and related fields accordingly:
ann_file='VOC2007/ImageSets/Main/train.txt',  
            img_subdir="PNGImages",
            ann_subdir="Annotations",
            data_prefix=dict(sub_data_root='VOC2007/'

Training

CUDA_VISIBLE_DEVICES=0 python tools/train.py configs/auxdet/auxdet_r50_fpn_1x_voc.py --work-dir your_path

Testing

CUDA_VISIBLE_DEVICES=0 python tools/test.py configs/auxdet/auxdet_r50_fpn_1x_voc.py "best.pth" --work-dir your_path

We provide our pretrained weight file here. Please download it and specify the path to "best.pth" in the command above.

---

⚖️ Benchmark

Category	Venue	Backbone	FLOPs ↓	Params. ↓	AP50 (%) ↑	Recall (%) ↑
General Object Detection
🔹 One-stage
RetinaNet	ICCV17	ResNet50	0.197T	36.330M	60.7	74.8
TOOD	ICCV19	ResNet50	0.191T	32.018M	53.2	61.9
FCOS	ICCV19	ResNet50	0.194T	32.113M	49.2	58.3
GFL	NeurIPS20	ResNet50	0.197T	32.258M	48.5	60.2
ATSS	CVPR20	ResNet50	0.194T	32.113M	50.1	60.6
AutoAssign	arXiv20	ResNet50	0.195T	36.244M	51.2	62.1
🔹 Two-stage
Faster R-CNN	NeurIPS15	ResNet50	0.200T	41.348M	31.2	33.5
Cascade R-CNN	TPAMI19	ResNet50	0.228T	69.152M	35.8	37.5
Grid R-CNN	CVPR19	ResNet50	0.313T	64.467M	32.2	34.8
Cascade RPN	NeurIPS19	ResNet50	0.190T	42.527M	74.4	84.6
Libra R-CNN	CVPR19	ResNet50	0.209T	41.611M	30.7	33.8
Dynamic R-CNN	ECCV20	ResNet50	0.200T	41.348M	35.0	37.2
SABL	ECCV20	ResNet50	0.198T	36.357M	64.1	77.8
🔹 End-to-End
Deformable DETR	ICLR21	ResNet50	0.189T	40.099M	33.1	49.8
Sparse R-CNN	CVPR21	ResNet50	0.146T	0.106G	40.4	62.1
DINO	arXiv22	ResNet50	0.268T	47.540M	67.1	82.8
DAB-DETR	ICLR22	ResNet50	99.511G	43.702M	17.6	31.9
Small Target Specialized
ALCNet	TGRS21	ResNet20	6.047G	42.700M	36.3	69.7
NWD	ISPRS22	DetectoRS (ResNet50)	0.235T	0.123G	48.5	53.6
RFLA	ECCV22	DetectoRS (ResNet50)	0.235T	0.123G	71.8	80.6
OSCAR	TGRS23	ResNet50	0.327T	30.730M	40.2	65.4
DNANet	TIP23	-	0.228T	4.697M	73.9	46.9
RDIAN	TGRS23	-	0.059T	0.217M	62.5	51.8
DQ-DETR	ECCV24	ResNet50	1.476T	58.680M	70.1	81.3
EFLNet	TGRS24	-	0.102T	38.336M	73.7	68.1
PConv	AAAI25	-	0.012T	2.922M	71.7	64.5
AuxDet (Ours)	-	ResNet50	0.215T	45.279M	77.9	87.2

🔖 Citation

---

@article{shi2025AuxDet,
	title={AuxDet: Auxiliary Metadata Matters for Omni-Domain Infrared Small Target Detection}, 
	author={Yangting Shi and Renjie He and Le Hui and Xiang Li and Jian Yang and Ming-Ming Cheng and Yimian Dai},
	year={2025},
	journal={arXiv preprint arXiv:2505.15184},
}