Dataset Hub

iHarmony4 is the first large‑scale image harmonization public benchmark dataset, containing four sub‑datasets: HCOCO, HAdobe5k, HFlickr, and Hday2night. Each sub‑dataset includes synthetic composite images, foreground masks of the composite images, and the corresponding real images.

CrowdHuman is a benchmark dataset for evaluating detector performance in crowded scenes. It is large‑scale, richly annotated, and highly diverse, comprising training, validation, and test sets with a total of 470,000 annotated human instances (average 23 persons per image) covering various occlusion conditions. Each person instance is annotated with a head bounding box, a visible region box, and a full‑body box.

A large‑scale ultra‑small video object detection benchmark dataset designed to support subsequent research and applications, providing annotations in COCO, COCOVID, and YOLO formats.