Few-Shot Object Detection via Variational Feature Aggregation

As few-shot object detectors are often trained with abundant base samples and fine-tuned on novel examples, the learned models usually biased to classes sensitive variance of examples. To address this issue, we propose a meta-learning framework two feature aggregation schemes. More precisely, first present Class-Agnostic Aggregation (CAA) method, where query support features can be aggregated regardless their categories. The interactions between different encourage class-agnostic representations reduce confusion classes. Based CAA, then Variational Feature (VFA) which encodes examples into class-level for robust aggregation. We use variational autoencoder estimate class distributions sample from that more Besides, decouple classification regression tasks so VFA is performed branch without affecting localization. Extensive experiments PASCAL VOC COCO demonstrate our method significantly outperforms strong baseline (up 16%) previous state-of-the-art methods (4% in average).