Few-shot object detection via semantic prompts and classifier decoupling.

Abstract

Many existing few-shot object detection methods employ two-stage detectors to achieve higher accuracy. Given the limited feature information and the challenges of adapting two-stage object detectors to few-shot learning, this paper proposes a few-shot object detection method based on semantic prompts and classifier decoupling. The key to incorporating textual information into object detectors lies in the effective fusion and alignment of image and text features. This paper introduces a Semantic Prompts module, enhancing the features of few-shot learning while aiding the model in better understanding image content. Leveraging the functionalities of the components of two-stage object detectors and their inter-component interactions, Gradient Scaling is employed to attenuate parameter updates, mitigating negative inter-module influences. To address the inconsistent feature demands between classification and regression branches, a Classifier Decoupling module is utilized to achieve more accurate classification and localization effects. Experimental evaluations on benchmark datasets demonstrate that the proposed method outperforms strong baselines such as DeFRCN by up to 3.15 % mAP under 1-shot settings on PASCAL VOC. These improvements indicate enhanced generalization and robustness in low-data regimes.