Gait3D is a large-scale gait recognition dataset based on 3D representation. The dataset is collected from 39 cameras in an unconstrained real indoor scene, containing 4,000 subjects and more than 25,000 sequences. In addition, the dataset provides 3D SMPL data recovered from video frames, which supports 3D modeling of dense human body shape, 3D perspective, and gait dynamic information. In addition, it also provides 2D human contour and human key point data to help researchers explore multimodal gait recognition.