Dual-Rate Dynamic Teacher for Source-Free Domain Adaptive Object Detection (ICCV 2025)

The core contribution is Asynchronous Exponential Moving Average (AEMA), which update teacher parameters in asynchronous manner.

The framework of the proposed Dual-rate Dynamic Teacher (DDT) is shown in the following figure.

Environment Setup

conda create -n ddt python=3.8
conda activate ddt
conda install pytorch==1.12.0 torchvision==0.13.0 torchaudio==0.12.0 cudatoolkit=11.3 -c pytorch
pip install -r requirements.txt

# Compiling and testing Deformable DETR CUDA operators
cd models/ops
bash make.sh
python test.py

Dataset

Our method used 3 popular SF-DAOD benchmarks:

• city2foggy: Cityscapes (source domain) → FoggyCityscapes with foggy level 0.02 (target domain).
• sim2city: Sim10k (source domain) → Cityscapes with car class (target domain).
• city2bdd: Cityscapes (source domain) → Bdd100k-daytime (target domain).

The raw data can be download from the official websites: Cityscapes, FoggyCityscapes, Sim10k, Bdd100k. The annotations are converted into COCO style, can download from here (provided by MRT-release). The datasets are organized as:

• data
  • cityscapes
    • annotations
    • leftImg8bit
      • train
      • val
  • foggy_cityscapes
    • annotations
    • leftImg8bit_foggy
      • train
      • val
  • bdd100k
    • annotations
    • images
      • 100k

Training & Evaluation

The source pretrained weight of DefDETR-R50 can be downloaded from DRU.

• outputs
  • city2foggy
    • source_only
      • downloaded weight
  • sim2city
    • source_only
      • downloaded weight
  • city2bdd
    • source_only
      • downloaded weight

Then, you can use the following shell to train and evaluate ddt.

bash configs/def-detr-base/city2foggy/teaching_mask_ddt.sh
bash configs/def-detr-base/city2foggy/evaluation_teaching_mask_ddt.sh

The training log file of city2foggy (best mAP = 45.9) is available at log file

Acknowledgement

This repository is based on DRU and MRT-release. We appreciate their great work.