FR-DETR

Official implementation for our paper accepted to ICME 2026.

This repository contains the implementation of FR-DETR, built on top of MMDetection, for object detection under degraded visual conditions.

The main implementation file is:

• mmdet/models/detectors/frdetr.py

Method

Configs

The main training configs are:

• config_rtdetr/frdetr_fog.py: FR-DETR on foggy VOC-style data.
• config_rtdetr/frdetr_dark.py: FR-DETR on low-light VOC-style data.

These configs are adapted from an RT-DETR-style structure, but instantiate the FRDETR detector.

Installation

This codebase follows the standard MMDetection installation flow. A Conda environment export is provided in environment.yml.

Tested core versions from the provided environment:

• Python 3.8.20
• PyTorch 2.0.1
• CUDA 11.8
• MMCV 2.0.1
• MMEngine 0.10.4

Create the environment and install the repo in development mode:

conda env create -f environment.yml
conda activate mmdet
pip install -e .

Datasets

For foggy and low-light VOC-style experiments, please prepare the datasets following Image-Adaptive-YOLO.

For adverse-weather experiments, please prepare the dataset following MODE.

The config files currently contain local absolute paths from our training environment. Before training or testing, update the following fields in each config:

• _base_
• pretrained
• data_root
• ann_file
• data_prefix

The expected data format follows MMDetection's VOCDataset format.

Training

Train FR-DETR on foggy data:

python tools/train.py config_rtdetr/frdetr_fog.py

Train FR-DETR on low-light data:

python tools/train.py config_rtdetr/frdetr_dark.py

Testing

After training, evaluate a checkpoint with:

python tools/test.py config_rtdetr/frdetr_fog.py work_dirs/frdetr_fog/best.pth

or:

python tools/test.py config_rtdetr/frdetr_dark.py work_dirs/frdetr_dark/best.pth

Citation

The citation for FR-DETR will be added after the arXiv or proceedings version is available.

This repository is based on MMDetection.