Spatial reasoning for few-shot object detection

Although modern object detectors rely heavily on a significant amount of training data, humans can easily detect novel objects using few examples. The mechanism the human visual system is to interpret spatial relationships among various and this process enables us exploit contextual information by considering co-occurrence objects. Thus, we propose reasoning framework that detects with only examples in context. We infer geometric relatedness between base RoIs (Region-of-Interests) enhance feature representation categories an detector well trained categories. employ graph convolutional network as their are defined nodes edges, respectively. Furthermore, present data augmentation overcome few-shot environment where all bounding boxes image resized randomly. Using PASCAL VOC MS COCO datasets, demonstrate proposed method significantly outperforms state-of-the-art methods verify its efficacy through extensive ablation studies.