Reproduce “Remedying Target-Domain Astigmatism for Cross-Domain Few-Shot Object Detection” 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
Yongwei Jiang, Yixiong Zou, Yuhua Li, Ruixuan Li
Researchers from Huazhong University of Science and Technology introduce a bio-inspired center-periphery attention refinement framework to remedy "target-domain Astigmatism" in Cross-Domain Few-Shot Object Detection. This framework, which includes Positive Pattern Refinement, Negative Context Modulation, and Textual Semantic Alignment, consistently achieves state-of-the-art performance, increasing average mAP from 29.36% to 33.73% in 5-shot settings across six diverse datasets.
Key insights
- Detection models exhibit "target-domain Astigmatism," a previously underexplored phenomenon where attention remains broadly dispersed in target domains compared to focused attention in source domains.
- The "attention distance" metric quantifies this dispersion, revealing that standard fine-tuning only marginally reduces this issue, highlighting the need for a targeted solution.
- Inspired by the human fovea-style visual system, refining attention by distinguishing between a focused "center" (foreground objects) and a broader "periphery" (background context) is crucial for effective cross-domain generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.