CrowdHuman is a large-scale, richly annotated human detection dataset. The dataset contains images collected from the Internet, with 15,000 images for training, 4,370 for validation, and 5,000 for testing. Compared with previous challenging pedestrian detection datasets such as CityPersons, the total number of images has increased by more than 10 times. In the CrowdHuman training subset, the total number of people is also significantly larger than other datasets, with 340,000 people and 90,000 ignore region annotations.

The dataset has a total of 470K human instances from training and validation subsets. Each image in the dataset contains about 23 people with various occlusions. Each human instance is annotated with head bounding-box, human visible-region bounding-box, and human full-body bounding-box.