Few-Shot Object Detection via Classify-Free RPN

The research community has shown great interest in few-shot object detection, which focuses on detecting novel objects with only a small number of annotated examples. Most of the works are based on the Faster R-CNN framework. However, due to the absence of annotated data for novel instances, models are prone to base class bias, which can result in misclassifying novel instances as background or base instances. Our analysis reveals that although the RPN is class-agnostic in form, the binary classification loss possesses class-awareness capabilities, which can lead to the base class bias issue. Therefore, we propose a simple yet effective classify-free RPN. We replace the binary classification loss of the RPN with Smooth L1 loss and adjust the ratio of positive and negative samples for computing the loss. This avoids treating anchors matched with novel instances as negative samples in loss calculation, thereby mitigating the base class bias issue. Without any additional computational cost or parameters, our method achieves significant improvements compared to other methods on the PASCAL VOC and MS-COCO benchmarks, establishing state-of-the-art performance.