Multi-View Part-Based Few-Shot Object Detection.

Abstract

Few-shot object detection (FSOD) aims to detect new object categories with a few annotated samples and has achieved great development. However, compared with general object detection, object misclassification remains a serious issue in FSOD. This is because, in scenarios with few samples, the model struggles to learn sufficient discriminative information to represent the corresponding classes. To address this issue, we propose a multi-view part-based FSOD network (MPFSOD), which extracts more discriminative information from multiple views to generate highly discriminative parts that effectively characterize objects, thereby promoting the accurate detection of new categories. Specifically, we first propose a part-based detector (PBD), where task-aware object-level parts are generated and used to enhance the discriminativeness of the object representations. Based on the PBD, we then introduce an image-level multi-view fusion module (Img-MVF) and an instance-level multi-view modulation module (Inst-MVM). These two modules extract richer discriminative information hidden in multiple views of the target, further facilitating the generation of discriminative parts in PBD. Extensive experiments on PASCAL VOC and MS COCO demonstrate that our method significantly outperforms a strong baseline (up to 11.2%) and previous state-of-the-art methods (4.3% in average).