Background suppression and comprehensive prototype pyramid distillation for few-shot object detection

Abstract

Few-Shot Object Detection (FSOD) methods can achieve detection of novel classes with only a small number of annotated samples and have received widespread attention in recent years. Meta-learning has been proven to be a key technology for addressing few-shot problems. Typically, meta-learning-based methods require an additional support branch to extract class prototypes for the few-shot classes, and the detection head performs classification and detection by measuring the distance between the class prototypes and the query features. Since the input to the support branch is the object image annotated with a bounding box, it often contains a large amount of background information, which degrades the quality of the class prototypes. Through our meticulous observation, we found that the center of the bounding box is often the core feature area of the object. Based on this, we designed a lightweight Background Suppression (BS) module that suppresses background features by measuring the similarity between the peripheral and central features of the support features, thereby providing high-quality support features for class prototype extraction. Additionally, in terms of class prototype extraction, we designed a more robust Comprehensive Prototype Pyramid Distillation (CPPD) module. This module first captures the multi-scale feature information of the object from the background-suppressed support features, and then uses a pyramid structure to hierarchically distill the multi-scale features to extract more comprehensive and purer class prototypes. Extensive experimental results on the PASCAL VOC and COCO datasets show that compared to other models under the same architecture and techniques, we achieved the best results.