Reproduce “Meta Faster R-CNN: Towards Accurate Few-Shot Object Detection with Attentive Feature Alignment” on OpenResearch

About this paper

Guangxing Han, Shiyuan Huang, Jiawei Ma, Yicheng He, Shih-Fu Chang

Meta Faster R-CNN introduces a meta-learning framework for few-shot object detection, addressing challenges in proposal generation and classification for novel categories through a Meta-RPN and an attentive feature alignment mechanism. The model achieves state-of-the-art performance across MSCOCO and PASCAL VOC benchmarks, demonstrating improved robustness particularly in extremely low-shot settings.

Key insights

• Metric-learning based non-linear classifiers within an RPN framework can significantly improve the recall and quality of object proposals for novel classes.
• Explicitly aligning features between noisy proposals and class prototypes, coupled with foreground attention, is crucial for accurate few-shot object classification despite spatial inaccuracies.
• A modular approach that separates base and novel class detection allows for robust performance on established categories while enabling flexible, incremental addition of new classes.

Summary by alphaXiv.