Reproduce “Enhancing Source-Free Domain Adaptive Object Detection with Low-confidence Pseudo Label Distillation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ilhoon Yoon, Hyeongjun Kwon, Jin Kim, Junyoung Park, Hyunsung Jang, Kwanghoon Sohn
A novel Low-confidence Pseudo Label Distillation (LPLD) method enhances Source-Free Domain Adaptive Object Detection (SFOD) by carefully mining and leveraging previously ignored low-confidence predictions corresponding to hard-to-detect objects. The approach, from Yonsei University and others, consistently improves performance across various benchmarks and significantly reduces false negatives for challenging object categories.
Key insights
- Low-confidence predictions, traditionally discarded as noise, contain valuable information about "hard positive" objects critical for robust domain adaptation, provided they are carefully mined and distilled.
- Employing a soft distillation approach (KL divergence) with adaptive weighting effectively leverages these noisy low-confidence pseudo-labels, avoiding direct hard supervision while still guiding the student model.
- Adaptive weighting, based on feature consistency between teacher and student, can prioritize more reliable low-confidence labels, enhancing optimization efficiency and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.