Enhancing Few-Shot Object Detection with Modified Faster R-CNN and Transfer Learning

Few-shot object detection, which aims to improve the generalization ability of the object detection model with limited data, has gained significant attention. Many existing methods use standard Faster R-CNN as the basic model and add modules to assist the detector to adapt to novel classes. However, Faster R-CNN as a method designed for the traditional object detection task, does not consider the few-shot condition and lacks relevant design for the few-shot detection. In this paper, we present an effective transfer learning-based approach for few-shot object detection. We propose modifications to the Faster R-CNN structure to address specific challenges in few-shot scenarios. Our method introduces a decoupled feature extraction network to obtain clearer and more generalized features, mitigating conflicts between classification and regression. Additionally, a feature complementing branch is introduced to obtain discriminative features for novel classes during fine-tuning. Furthermore, we design a margin cosine loss to enhance the classifier’s performance. Compared with baseline, our method achieves an average improvement of +7.0 nAP50 on PASCAL VOC and +3.0 nAP on COCO. Experimental results show that our method can outperform many existing algorithms with the state-of-the-art performance achieved.