Reproduce “Few-Shot Object Detection via Variational Feature Aggregation” on OpenResearch

About this paper

Jiaming Han, Yuqiang Ren, Jian Ding, Ke Yan, Gui-Song Xia

This research from Wuhan University and Tencent YouTu Lab introduces Variational Feature Aggregation (VFA) for few-shot object detection, which models class distributions using VAEs to generate robust support features. The method, built on an enhanced Meta R-CNN++ baseline, significantly improves performance and robustness, particularly in low-shot settings, outperforming prior state-of-the-art on PASCAL VOC by a substantial margin.

Key insights

• Parameter freezing, cosine classifiers, and strategic last-layer initialization are crucial for building highly competitive meta-learning baselines in FSOD.
• Modeling class-level features as probability distributions via VAEs effectively mitigates the impact of high variance in few-shot examples, leading to more robust class prototypes.
• Decoupling classification and regression tasks in the detection head, where classification uses aggregated support features and regression uses original query features, significantly improves overall detection performance by optimizing features for each specific requirement.

Summary by alphaXiv.