NBA2K Dataset for the ECCV2020 paper : Reconstructing NBA Players
This repository maintains the NBA2K dataset of our ECCV 2020 paper 'Reconstructing NBA Players'. We collect RGB images, triangle meshes with UV coordinates, texture maps, 3D human poses and camera projection matrix from the NBA2K19 game engine. Our dataset has large diversity in basketball poses and provides industry-level clothed basketball player meshes.
/release/${player_name}/${capture_mode}/${capture_name}.png
: RGB images of frame captures./release/${player_name}/${capture_mode}/${capture_name}/players/${person_id}_person_v2_transform_v2.npy
: 3D bone transformation of players./release/${player_name}/${capture_mode}/${capture_name}/proj_mat.npy
: 4x4 camera projection matrix.rest_pose_data/
: 0_person_no_dup.obj
stores A-pose coarse mesh. 0_person_no_dup_blend.npy
stores bone id matrix for A-pose coarse mesh, used for Linear Blend Skinning. aug_joints_v2.obj
stores A-pose 3D joints. aug_joints_v2.npy
stores bone id matrix for A-pose 3D joints. aug_joints_v2_id.txt
stores the joint definition.release/${player_name}/${capture_mode}/${capture_name}/player_parts/
: Raw triangle meshes of different body parts for all players in current frame.release/${player_name}/${capture_mode}/${capture_name}/players/
: ${person_id}_person.obj
stores the full body mesh (including hair, tooth, beard, mane, eye etc.) of all players in current frame. ${person_id}_person_simple.obj
stores the simplified full body mesh (only has shirt, pant, leg, arm, head, shoes) of all players in current frame. These meshes are not registered and have different mesh topologies.release/${player_name}/resampled/${capture_mode}/${capture_name}/player_parts/
: Registered triangle meshes of different body parts for all players in current frame. Only has shirt, pant, leg, arm, head, shoes.release/${player_name}/resampled/${capture_mode}/${capture_name}/players/
: Registered simplified full body meshes for all players in current frame. Only has shirt, pant, leg, arm, head, shoes. The mesh topology is same across all data examples and can be used for deep network training.release/${player_name}/textures
: Texture maps of the current player. To make sure the mtl file can find the texture images, you need to copy the textures folder under the ${capture_mode} sub directory. Otherwise, you will encounter texture not found error when you visualize the obj file in Meshlab.seg_release/seg_${player_name}_${capture_mode}.txt
: Body segmentation file. It indicates which body part belongs to which player.Processed data used to train the mesh generation networks.
template/objs/
: Rest pose mesh for all players in obj format.template/template.pkl
: A dictionary with key=player_name and value=vertices of the rest pose mesh.release/${player_name}/images
: Center cropped 256x256 images for target players.release/${player_name}/release_${player_name}_${capture_mode}.pkl
: Meta data dictionary with following keys:
Matrix mapping from vertices position to joints position.
You can directly use the training data in mesh_release.zip. We also provide scripts for processing the raw data. Here are the necessary steps to run the script:
unzip images.zip -d images
.root_folder
in prepare_data.py to your DATA_ROOT_FOLDER, then run python prepare_data.py
.root_folder
in gen_line_maps.py to your DATA_ROOT_FOLDER, then run python gen_line_maps.py
.Please fill in the Google form to request the dataset.
As we said in the paper, we are not allowed to release the data of current NBA players due to copyright issues. Instead, we additionally collected the same kind of data for 28 synthetic players. The original data are captured from real NBA players under Lakers' 2018-19 Home/Away uniforms and Nike LeBron 16 shoes. The released data are captured from 2K-made synthetic players under Raptors' 1946-47 Home uniforms and 2K brand shoes. The number of vertices and faces (V_num, F_num) for original data and released data are as follows:
We thank Visual Concepts for allowing us to capture, process, and share our extracted NBA2K data for research.
The dataset is made available under Creative Commons BY-NC-SA 4.0 license by University of Washington. You can use, redistribute, and adapt it for non-commercial purposes, as long as you (a) give appropriate credit by citing our paper, (b) indicate any changes that you've made, and (c) distribute any derivative works under the same license.
If you use our dataset, please citing our work.
@InProceedings{zhu_2020_eccv_nba,
author={Zhu, Luyang and Rematas, Konstantinos and Curless, Brian and Seitz, Steve and Kemelmacher-Shlizerman, Ira},
title={Reconstructing NBA players},
booktitle = {Proceedings of the European Conference on Computer Vision (ECCV)},
month = {August},
year={2020}
}