Abstract Few-shot object detection aims to accurately detect novel classes using a limited number of sample instances. Currently, two-stage detection methods in the fine-tuning paradigm have been widely recognized as the main strategy for FSOD. However, during the fine-tuning process, the addition of novel class representations often causes shifts and distortion in the feature distribution of base classes due to knowledge transfer. As a result, it is challenging to mitigate catastrophic forgetting for base classes while ensuring improvements in the performance of novel classes. To solve this problem, we focus on fine-tuning stage of two-stage paradigm and propose multi-dimensional feature adaptive calibration (MDFAC) method. Specifically, we propose Equiangular Tight Frame Guidance Module (ETFGM) to construct a high-dimensional hypersphere memory bank to store the pre-trained base class distributions. This module guides the classifier to follow the uniform distribution of each class center, combating catastrophic forgetting of base class knowledge and ensuring independent learning of novel class knowledge. Meanwhile, adaptive calibration classification (ACC) loss dynamically adjusts the model’s attention, prioritizing categories that are perceived less favorably based on the real-time detection frequency of the classifier. Through the synergistic integration of ETFGM and ACC loss, the classifier is autonomously trained to improve its discriminative ability. Extensive benchmark results on PASCAL VOC and MS COCO datasets demonstrate that our method improves the average detection performance for novel classes by 1.8% (nAP50) and 2.1% (nAP), respectively, while maintaining competitive performance on base classes compared to existing methods. Overall, our approach outperforms the state-of-the-art.
Xie et al. (Thu,) studied this question.