NBA2K Dataset

Overview

This dataset contains RGB images, triangle meshes (with UV coordinates), texture maps, 3D human poses, and camera projection matrices extracted from the NBA2K19 game engine. It offers high diversity in basketball postures and provides industry‑level meshes of clothed basketball players.

Data Content

Capture Modes:

• 2ku: Collected in the 2KU freestyle mode; each frame contains a single player, primarily for dribbling, layup, dunk, and shooting postures.
• normal: Collected during fast‑match mode; each frame may contain up to 10 players, including occlusions and various normal in‑game postures.
• rest pose: Collected in player‑editing mode; each player’s standing pose is captured once.

File Structure

• images.zip: Contains /release/${player_name}/${capture_mode}/${capture_name}.png – the RGB frame captures.
• pose.zip: Contains /release/${player_name}/${capture_mode}/${capture_name}/players/${person_id}_person_v2_transform_v2.npy (the 3D skeletal transforms) and /release/${player_name}/${capture_mode}/${capture_name}/proj_mat.npy (the 4×4 camera projection matrix).
• mesh.zip: Includes raw and registered triangle meshes, texture maps, and body segmentation files.
• mesh_release.zip: Pre‑processed data for training mesh‑generation networks.

Data Processing

Processing scripts are provided; they require numpy, opencv‑python, shapely, and imgaug and sufficient disk space (over 100 GB).

Data Request

Please fill out the Google Form to request the dataset.

Important Note

Due to copyright restrictions, real NBA player data cannot be released; instead, data for 28 synthetic players are provided.

License

The dataset is offered under the Creative Commons BY‑NC‑SA 4.0 license, permitting non‑commercial use, redistribution, and adaptation, provided attribution is given and derivative works are shared publicly.

Citation

If you use this dataset, please cite the related paper.