Dynamic Routing and Calibration for Few-Shot Object Detection

Abstract

Few-shot object detection (FSOD), aiming to enhance the performance of novel object detection with limited labeled samples, has recently gained significant attention. Recent researches primarily focus on improving the generalization of novel classes and enhancing detector performance. However, the diversity of samples is often overlooked, and object proposals with inaccurate classifications or locations remain uncorrected. In this paper, we propose Dynamic Routing and Calibration for Few-Shot Object Detection (DRC-FSOD). Our approach includes a dynamic backbone routing that adapts to various samples by selecting appropriate backbones dynamically. Meanwhile, we construct a dynamic calibration module, which dynamically perform individual calibration for proposals based on their scores. Experimental results on MS COCO and Pascal VOC datasets show superiority over state-of-the-art methods.