TotalCapture is a large-scale human pose estimation dataset. The dataset is recorded indoors by 8 20HZ high-definition cameras, with a total of 5 volunteers, including 4 men and 1 woman. Each volunteer performed 4 different actions: moving joints, walking, action, and freestyle, and repeated these actions 3 times.

The dataset has publicly released foreground matte and RGB images, and the ground truth pose is obtained through a marker-based motion capture system with markers smaller than 5mm.