Few-Shot Object Detection with Enhanced Prototypes

Few-shot object detection (FSOD) seeks to detect novel classes with limited training instances and has received a great deal of interest. Existing meta-learning based methods extract prototypes to aggregate with query features for instance classification and localization. However, they often suffer from lack of effective prototypes. In this paper, we propose a FSOD model, termed enhanced prototype net (EPNet), based on meta-learning. We first present an early stage aggregation module, which aggregates cross-attention query features and support features to enhance prototypes generating. Then a feature convex combination mechanism is proposed to perform data augmentation in the feature space and a soft distance loss to shrink the regions of prototypes features, thus extracting distinct and robust prototypes. Besides, a class-agnostic aggregation module is developed to fuse the query features with every prototype to boost training query samples for instance classification. Experimental results on Pascal VOC and MS COCO datasets demonstrate that the proposed model surpassed the baseline model and achieved state-of-the-art performances.