Few-Shot Object Detection via Sample Processing

Few-shot object detection (FSOD) eliminates the dependence on tremendous instances with manual annotations in conventional detection. We deem that scarcity of positive samples is main reason restricts performance FSOD detectors. In this paper, a novel model via sample processing, namely, FSSP, proposed to detect objects accurately only few annotated samples, which based structural design Siamese network and uses YOLOv3-SPP as baseline. Central FSSP are our designed self-attention (SAM) positive-sample augmentation (PSA) modules. The former attempts better extract representative features hard latter expands number enriches scale distribution inhibiting growth negative samples. For fine-tuning phase, we modify classification loss function increase punishment for Experiments conducted PASCAL VOC MS COCO datasets confirm achieves competitive compared state-of-the-art