Instance-level feature representation calibration for visual object detection

Abstract

Few-shot object detection has gained significant attention due to the scarcity of training samples in real-world applications. Most existing methods attempt to transfer knowledge learned from abundant base classes to novel class detection, typically following a two-stage process: base training and fine-tuning. While these detectors excel at object localization, they often struggle with classification due to biased feature representations of instances. In this paper, we propose a novel framework for feature representation learning in few-shot object detection, aimed at refining the instance representation using a prototype-based supervised contrastive learning approach. Specifically, we design a prototype representation bank that serves as a template for supervised contrastive learning and introduce a positive example learning strategy to obtain generalized and discriminative object features. Additionally, we introduce a balanced cross-entropy loss that dynamically adjusts the weightings of gradients from positive and negative samples, thereby enhancing the confidence in object recognition. Extensive experiments on the Pascal VOC and MS-COCO benchmarks show that our method achieves state-of-the-art performance, with significant improvements across most splits and shot settings.