Reproduce “Reference-based Category Discovery: Unsupervised Object Detection with Category Awareness” 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
Yichen Li, Qiankun Liu, Ying Fu
Researchers from Beijing Institute of Technology and University of Science and Technology Beijing introduced Reference-based Category Discovery (RefCD), an unsupervised method for object detection that provides category awareness by matching object features to unlabeled reference images. This approach achieved 15.1% AP on COCO NOVEL, outperforming several supervised one-shot detection methods, and demonstrated competitive category-agnostic detection.
Key insights
- Category awareness can be learned in a fully unsupervised manner by leveraging feature similarity between detected objects and dynamic reference images, guided by a specialized loss function.
- The Feature Similarity (FS) loss is crucial for aligning the detector's internal query features with reference features, effectively bridging the gap between unsupervised training and category-aware inference decisions.
- Utilizing robust class token features from a self-supervised Vision Transformer (DINOv2) as reference embeddings provides rich semantic information, enhancing both category-aware and category-agnostic detection performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.