Focus the Overlapping Problem on Few-Shot Object Detection via Multiple Predictions

Mandan Guan, Wenqing Yu, Yurong Guo, Keyan Huang, Jiaxun Zhang, Kongming Liang, Zhanyu Ma

Abstract: The overlapping problem is one of the main challenges of few-shot object detection (FSOD). It refers to the situation where some parts of the target are obscured by other objects or occluders. Since the support set is insufficient to provide enough samples with overlapping objects, it is difficult to get the correct bounding boxes due to the missing information. In this paper, we aim to detect highly overlapping instances in few-shot scenes and present a proposal-based method by combining both basic training and fine-tuning stages. Specifically, we predict a set of bounding boxes instead of a single bounding box for each proposal after proposal generation. Then, we introduce a new NMS strategy to prevent the erroneous removal of our desired bounding box by traditional NMS. Simultaneously, we introduce a new loss to distinguish the obtained bounding boxes to avoid the convergence of bounding boxes during the final training process. While benchmarking on MS-COCO and Pascal VOC, our method is confirmed to be efficient and obtains good generalizability. Our proposed method served as a good stimulus for future study in the field of FSOD.