Bi-path Combination YOLO for Real-time Few-shot Object Detection

Abstract

Few-shot object detection (FSOD) has more attention in recent years as the quantitative limitation of instances during the model training. Previous works based on meta learning and transfer learning focus on the detection precision but ignore the inferring speed, which is difficult to apply in amounts of applications. In this letter, to keep a high inferring speed and a comparable detection precision, we propose a real-time detector entitled Bi-path Combination You Only Look Once (BC-YOLO) for FSOD. BC-YOLO can be categorized as a transfer learning based one-stage object detector with a two-phase training scheme. It is particularly composed of bi-path parallel detection branches which detect base and novel class objects respectively and commonly detect objects with a discriminator in the inferring stage. Moreover, to elevate the model generalization trained from few-shot objects, we further propose an Attentive DropBlock algorithm to make the detector focus on the entire details of objects instead of the local discriminative regions. Extensive experiments on PASCAL VOC 2007 and MS COCO 2014 datasets demonstrate that our method can achieve a better tradeoff between speed and precision than state-of-the-art methods.