COCONut (COCO Next Universal segmenTation dataset) is a large-scale panoramic image segmentation dataset released by ByteDance.This dataset is the first large-scale manually annotated segmentation dataset, containing about 383K images and 5.18 million manually annotated panoramic segmentation masks. This achievement has been selected for CVPR2024.

The release of COCONut greatly improves the community's ability to evaluate the progress of new neural networks. The dataset is semantically rich and the mask segmentation is fine-grained. The dataset also supports a variety of understanding tasks, such as panoptic segmentation, instance segmentation, semantic segmentation, object detection, semantic control generation, and open vocabulary segmentation. On many tasks, significant performance improvements are achieved simply by replacing the dataset.