# SSDA-YOLO: Semi-supervised Domain Adaptive YOLO for Cross-Domain Object Detection

**Authors:** Huayi Zhou, Fei Jiang, Hongtao Lu  
**arXiv:** [2211.02213](https://arxiv.org/abs/2211.02213)  
**Date:** 2022-11-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2211.02213v1/

## Key points

- Introduction of Semi-supervised Domain Adaptive YOLO (SSDA-YOLO) for domain adaptive object detection (DAOD)
- Existing DAOD methods rely on computationally intensive two-stage detectors
- Integration of compact one-stage detector YOLOv5 with domain adaptation techniques
- Utilization of knowledge distillation framework with Mean Teacher model for obtaining instance-level features from unlabeled data in target domain
- Employment of scene style transfer to generate pseudo images in different domains
- Introduction of intuitive consistency loss to align cross-domain predictions effectively
- Evaluation on public benchmarks such as PascalVOC, Clipart1k, Cityscapes, and Foggy Cityscapes
- Verification of generalization capability through experiments on yawning detection datasets from various classrooms
- Significant improvements in DAOD tasks using SSDA-YOLO method
- Availability of code implementation on GitHub at \url{https://github.com/hnuzhy/SSDA-YOLO}
- Promising approach combining compact one-stage detector with effective domain adaptation techniques to enhance cross-domain object detection performance significantly

## Summary

This paper introduces a novel approach called Semi-supervised Domain Adaptive YOLO (SSDA-YOLO) for domain adaptive object detection (DAOD). DAOD aims to address the performance degradation caused by differences between source and target domains. However, most existing DAOD methods rely on computationally intensive two-stage detectors, which are not ideal for industrial applications. To overcome this limitation, the authors propose integrating the compact one-stage detector YOLOv5 with domain adaptation techniques. They utilize the knowledge distillation framework with the Mean Teacher model to assist the student model in obtaining instance-level features from unlabeled data in the target domain. Additionally, they employ scene style transfer to generate pseudo images in different domains, addressing image-level differences. Furthermore, the authors introduce an intuitive consistency loss to align cross-domain predictions more effectively. The proposed SSDA-YOLO method is evaluated on public benchmarks such as PascalVOC, Clipart1k, Cityscapes and Foggy Cityscapes. To verify its generalization capability experiments are conducted on yawning detection datasets collected from various classrooms. The results demonstrate significant improvements in DAOD tasks using SSDA-YOLO. The code for implementing this method is made available on GitHub at \url{https://github.com/hnuzhy/SSDA-YOLO}. Overall, this paper presents a promising approach that combines a compact one-stage detector with effective domain adaptation techniques to enhance cross-domain object detection performance significantly.