Large-scale datasets such as ImageNet, PASCAL VOC, and COCO play important roles in the recent success of deep learning algorithms in image recognition tasks. However, there are not sufficient datasets specifically designed for few-shot learning, especially in the few-shot semantic segmentation domain. We build the first large-scale few-shot segmentation dataset, FSS-1000, which consists of 1000 object classes with pixelwise annotation of ground-truth segmentation. Unique in FSS-1000, our dataset contains a significant number of objects that have never been seen or annotated in previous datasets, such as tiny daily objects, merchandise, cartoon characters, logos, etc.

We adapt the structure of Relation Network to build our baseline few-shot segmentation model to validate FSS-10...[ Read more ]

Large-scale datasets such as ImageNet, PASCAL VOC, and COCO play important roles in the recent success of deep learning algorithms in image recognition tasks. However, there are not sufficient datasets specifically designed for few-shot learning, especially in the few-shot semantic segmentation domain. We build the first large-scale few-shot segmentation dataset, FSS-1000, which consists of 1000 object classes with pixelwise annotation of ground-truth segmentation. Unique in FSS-1000, our dataset contains a significant number of objects that have never been seen or annotated in previous datasets, such as tiny daily objects, merchandise, cartoon characters, logos, etc.

We adapt the structure of Relation Network to build our baseline few-shot segmentation model to validate FSS-1000. By adopting networks such as VGG-16, ResNet-101, and Inception as backbones, we found that training our model from scratch using FSS-1000 achieves competitive and even better results than training with weights pre-trained by ImageNet which is more than 100 times larger than FSS-1000. Both our approach and dataset are simple, effective, and extensible to learn the segmentation of new object classes given very few annotated training examples.