Reproduce “Proposal Refinement 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
Yuan Zeng, Bin Song, Jie Guo, Yuwen Chen
A research team from Xidian University introduces a proposal refinement method for few-shot object detection, tackling the imbalance of region proposals between base and novel classes. This approach employs a refinement loss during base training and an RPN refinement branch during fine-tuning, resulting in consistent gains of 1-6% mAP50 for novel classes on PASCAL VOC and up to 1.8% [email protected] on MS COCO without increasing inference time.
Key insights
- The inherent insensitivity of feature extraction backbones, when pre-trained on abundant base classes, significantly contributes to the under-representation of novel class proposals in few-shot detection.
- Mitigating the impact of 'discouraging gradients' from base classes during initial training is crucial for a model to retain sensitivity and learn effective features for novel categories.
- Directly manipulating the proposal generation mechanism in the RPN by incorporating novel vs. base class awareness can effectively rebalance proposal distribution, leading to improved novel class detection without adding inference overhead.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.