Summary of Simplifying Source-free Domain Adaptation For Object Detection: Effective Self-training Strategies and Performance Insights, by Yan Hao et al.
Simplifying Source-Free Domain Adaptation for Object Detection: Effective Self-Training Strategies and Performance Insights
by Yan Hao, Florent Forest, Olga Fink
First submitted to arxiv on: 10 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Machine Learning (cs.LG)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary The paper explores source-free domain adaptation for object detection in computer vision, a challenging task with significant practical implications. Recent approaches have employed teacher-student architectures with various feature alignment and regularization strategies to address this issue. In contrast, the authors investigate simpler methods and their performance compared to more complex techniques in several scenarios. Notably, they highlight the importance of batch normalization layers in the detector backbone and demonstrate that adapting only these statistics is a strong baseline for source-free object detection (SFOD). The authors also propose the Source-Free Unbiased Teacher (SF-UT) method, which outperforms many previous SFOD approaches. Additionally, they show that training on a fixed set of pseudo-labels can achieve similar performance to more complex teacher-student mutual learning while being computationally efficient and mitigating teacher-student collapse. Experiments are conducted on several driving datasets, including Cityscapes, Sim10k, and KITTI, achieving a notable improvement of 4.7% AP50 on Cityscapes→Foggy-Cityscapes compared to the latest state-of-the-art in SFOD. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary The paper looks at how to make object detection work well on new images even if those images haven’t been labeled before. This is a tough problem because it’s hard to get lots of labeled data for every new situation. The authors try some simpler ideas and compare them to more complex methods that have been proposed recently. They find that just adjusting the way the detector looks at the images can be a good starting point, and they also propose a new method called Source-Free Unbiased Teacher (SF-UT) that works well. Another simple approach they try is training the detector on some fake labels, which surprisingly works almost as well as the more complex methods. |
Keywords
* Artificial intelligence * Alignment * Batch normalization * Domain adaptation * Object detection * Regularization
