VOccl3D is a large-scale synthetic dataset released by the University of California in 2025, focusing on 3D human understanding in complex occluded scenes. The related paper is titled "VOccl3D: A Video Benchmark Dataset for 3D Human Pose and Shape Estimation under real OcclusionsThe goal is to provide a more realistic evaluation benchmark for human pose estimation, reconstruction, and multimodal perception tasks, based on occlusion conditions.

This dataset contains over 250,000 images and approximately 400 video sequences, constructed from background scenes, human actions, and diverse textures, specifically including:

• Background: 40 real-world 3D representations learned from DL3DV, including various types of natural occlusion.
• Human motion: Approximately 400 motion sequences from AMASS
• Body textures: Approximately 200 textures from SMPLitex, covering different clothing, skin tones, and body types.

All sequences are rendered at 720×720 resolution and 30 fps, providing accurate camera intrinsic and extrinsic parameters. The dataset also provides multimodal annotations, including 3D pose and shape, 2D keypoints, human contours, semantic segmentation, occlusion labels, and human bounding boxes, which can be used to study multi-task human perception capabilities under occlusion conditions.