A Closer Look at Few-Shot Object Detection

Few-shot object detection, which aims to detect unseen classes in data-scarce scenarios, remains a challenging task. Most existing works adopt Faster RCNN as the basic framework and employ fine-tuning paradigm to tackle this problem. However, the intrinsic concept drift in the Region Proposal Network and the rejection of false positive region proposals hinder model performance. In this paper, we introduce a simple and effective task adapter in RPN, which decouples it from the backbone network to obtain category-agnostic knowledge. In the last two layers of the task adapter, we use large-kernel spatially separable convolution to adaptively detect objects at different scales. In addition, We design an offline structural reparameterization approach to better initialize box classifiers by constructing an augmented dataset to learn initial novel prototypes and explicitly incorporating priors from base training in extremely low-shot scenarios. Extensive experiments on various benchmarks have demonstrated that our proposed method is significantly superior to other methods and is comparative with state-of-the-art performance.