FoxInst: A Frustratingly Simple Baseline for Weakly Few-shot Instance Segmentation

We propose the first weakly-supervised few-shot instance segmentation task and a frustratingly simple but strong baseline model, FoxInst. Our work is distinguished from other approaches in that our method is trained with weak annotations, i.e., class and box annotations, during all phases, which leads to further data efficiency and practicality. Considering the challenging regime of our problem, we design the network to be an anchor-free architecture to avoid anchor box restriction, and train the network in a simple and stable way that first trains the whole network on the base classes, and then only fine-tunes the heads partially with few novel class data. To establish the foundation as a strong baseline, we carefully design evaluation setups by correcting the existing problems in the evaluation metric and test set, so that the effects of each component are well revealed. We show that FoxInst achieves comparable or even higher performance with the prior fully-supervised FSIS networks on COCO and PASCAL VOC datasets. We will release the code if accepted for reproduction.