Abstract:
Few-shot object detection, which aims to recognize unseen objects with a few annotated instances, has attracted increasing attention in the computer vision community. Most recent works tackle this problem under the meta-learning framework based on an episodic training strategy. In this work, we advance the few-shot object detection paradigm towards a new scenario called semi-supervised few-shot object detection (SSFSOD), where the unlabeled data are available within each episode. To address this paradigm, we propose a novel method which utilizes a dual model (teacher-student) to leverage available unlabeled data. Specifically, the teacher model provides high-quality pseudo-labels for the student model during the training process, while the student model uses the exponential moving average strategy to update the teacher model online. We also employ a two-fold correlation-guided attention module to guide RPN to generate task-specific region proposals by highlighting potential regions and informative channels. We conduct extensive experiments on three datasets MS COCO, PASCAL VOC, and FSOD. The experimental results demonstrate the effectiveness of the proposed method.
