"GTA-MMP"

"We built the GTA-MMP dataset based on the open-world game GTA-V. This dataset is a synthetic human action dataset with fixed and motion perspectives, encompassing daily and high-risk actions. It provides complete 2D&3D joint coordinates, environmental information, joint-by-joint occlusion labels, and camera parameters for each frame. It aims to provide a multi-view and high-quality annotation platform for pose estimation and complex action recognition. By directly accessing the game engine's interface, we automatically generate the following complete ground truth labels for each frame captured:Video and image data. High-resolution RGB images and continuous video streams are saved. Video and image sizes are 1920\u00d71080, and the frame rate is 60 fps.2D and 3D human body information. Includes precise 2D image coordinates of the character's 55 body joints and their 3D coordinates in the game world coordinate system.Occlusion labels. Provides visibility labels for each joint, including ``visible\", ``occluded\", and ``self-occluded\".Camera parameters. Provides the camera's 3D coordinates, vertical field of view, pitch angle, yaw angle, and roll angle in the game world coordinate system for each frame. This enables further calculation of the camera's intrinsic parameters (focal length, principal point, etc.) and extrinsic parameters (rotation matrix, translation matrix, etc.).Environmental information. Provides the character's gender, weather, and time as set during data collection."