Reproduce “Fine-Grained Prototypes Distillation for 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

Zichen Wang, Bo Yang, Haonan Yue, Zhenghao Ma

FPD introduces a method for few-shot object detection that distills fine-grained prototypes from mid-level features and enhances high-level feature fusion. It achieved state-of-the-art results on PASCAL VOC, notably improving performance by up to 4.0 percentage points AP50 on some novel sets, and competitive scores on MS COCO.

Key insights

  • Leveraging fine-grained prototypes, distilled from mid-level features, is crucial for capturing subtle local context and improving object detection accuracy in few-shot scenarios.
  • Distilling representative features with learnable queries offers a more efficient and robust alternative to direct dense feature matching for modeling detailed relationships.
  • A balanced approach to high-level feature aggregation, incorporating both positive and negative class samples through non-linear fusion, enhances the model's ability to discriminate between classes.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.