Summary of Enhancing Source-free Domain Adaptive Object Detection with Low-confidence Pseudo Label Distillation, by Ilhoon Yoon et al.
Enhancing Source-Free Domain Adaptive Object Detection with Low-confidence Pseudo Label Distillation
by Ilhoon Yoon, Hyeongjun Kwon, Jin Kim, Junyoung Park, Hyunsung Jang, Kwanghoon Sohn
First submitted to arxiv on: 18 Jul 2024
Categories
- Main: Computer Vision and Pattern Recognition (cs.CV)
- Secondary: Artificial Intelligence (cs.AI)
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 This paper introduces Low-confidence Pseudo Label Distillation (LPLD) loss, a novel approach for Source-Free domain adaptive Object Detection (SFOD). SFOD aims to deploy trained detectors to new domains without accessing source data, addressing concerns around data privacy and efficiency. Most SFOD methods rely on Mean-Teacher (MT) self-training with High-confidence Pseudo Labels (HPL), but these HPL often overlook small instances or ignore instances with low confidence. LPLD leverages proposals from Region Proposal Network (RPN) to extract Low-confidence Pseudo Labels (LPL), refining them using class-relation information and feature distance. The method outperforms previous SFOD methods on four cross-domain object detection benchmarks, effectively adapting by reducing false negatives and facilitating domain-invariant knowledge transfer. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary This paper helps us adapt object detectors to new places without needing the original data. This is important because it protects people’s privacy and makes things more efficient. Right now, most detector-adapting methods use a type of self-training that relies on very confident predictions. But these predictions often miss small objects or ignore ones that are tricky to detect. The new method in this paper uses proposals from the Region Proposal Network to find harder-to-detect objects, and then refines those predictions using information about how different object classes relate to each other. It also adjusts its focus based on how similar the objects look. This approach does a better job than existing methods of adapting detectors to new places. |
Keywords
» Artificial intelligence » Distillation » Object detection » Region proposal » Self training
