Introduction

This is the official repository for the paper Versatile Teacher: A Class-aware Teacher-student Framework for Cross-domain Adaptation.

Preparation

Requirements

This repository is built upon YOLOv5 and follow the same requirements. Please refer to the original repository for more details.

Datasets

Prepare your own datasets in YOLO format and create a corresponding .yaml file in the folder data.

Train

Firstly, use the script

python train.py \
    --cfg models/yolov5l_da.yaml \
    --weights weights/yolov5l.pt \
    --data data/src_data.yaml \
    --data data/tgt_data.yaml \
    --device 0 \
    --epochs 50 \
    --batch-size 16 \

to initialize the teacher model, saved as weights/yolov5l_teacher.pt.

Then, use the script

python train.py \
    --cfg models/yolov5l_da.yaml \
    --cfg models/yolov5l.yaml \
    --weights weights/yolov5l_teacher.pt \
    --weights weights/yolov5l_teacher.pt \
    --data data/src_data.yaml \
    --data data/tgt_data.yaml \
    --device 0 \
    --epochs 50 \
    --batch-size 16 \

to train the student model.

Citation

@article{YANG2025111024,
title = {Versatile Teacher: A class-aware teacher–student framework for cross-domain adaptation},
journal = {Pattern Recognition},
volume = {158},
pages = {111024},
year = {2025},
issn = {0031-3203},
doi = {https://doi.org/10.1016/j.patcog.2024.111024},
url = {https://www.sciencedirect.com/science/article/pii/S0031320324007751},
author = {Runou Yang and Tian Tian and Jinwen Tian},
}