Reproduce “DiGeo: Discriminative Geometry-Aware Learning for Generalized Few-Shot Object Detection” on OpenResearch

About this paper

Jiawei Ma, Yulei Niu, Jincheng Xu, Shiyuan Huang, Guangxing Han, Shih-Fu Chang

DiGeo, a framework from Columbia University, enhances generalized few-shot object detection by explicitly learning discriminative geometry-aware features. It improves novel-class detection performance while preserving base-class accuracy by fostering inter-class separation and intra-class compactness in the feature space, outperforming state-of-the-art methods on PASCAL VOC and MS COCO.

Key insights

• Explicitly managing the geometric properties of feature distributions, specifically inter-class separation and intra-class compactness, is crucial for robust generalized few-shot object detection.
• Utilizing a stable, pre-defined Simplex ETF for classifier weights (offline solution) is more effective and stable than online regularization for achieving maximal inter-class separation.
• Adaptive, class-specific margins, combined with novel instance up-sampling, are vital for efficient intra-class compactness and for balancing learning across highly imbalanced base and novel classes.

Summary by alphaXiv.