COCONut: A High-Quality, Large-Scale Dataset for Next-Gen Segmentation Models

COCONut is a novel, large-scale universal segmentation dataset proposed to modernize COCO. It features human-verified mask labels for 383K images. The dataset is created through an innovative assisted-manual annotation pipeline leveraging neural networks. The size of COCONut varies from COCONut-S (118K images) to COCONut-L (358K images with 4.75M masks).