ember-lab-berkeley/LATTE-MV

Dataset format:-

release_data/
├── match1/
│   ├── match1_3/
│   │   ├── match1_3.mp4               # Original clipped video
│   │   ├── match1_3_ball.csv          # Ball tracking results
│   │   ├── match1_3_keypoints_2d.npy  # 2D keypoints for detected humans
│   │   ├── match1_3_keypoints_3d.npy  # 3D keypoints for detected humans
│   │   ├── match1_3_metadata.json     # Metadata for detected humans
│   │   ├── match1_3_paddle.csv        # Paddle detection results
│   │   ├── match1_3_table.npy         # Detected table corners
│   │   ├── match1_3_recon.npy         # Reconstructed scene information
│   ├── ...
└── ...

NOTE: match{match_id}_{rally_id}_recon.npy consists of a 3d numpy array of shape (no of frames, 91, 3) with each row containing scene information for a frame in that timestep. The columns are as follows:-

• metadata ((fps, no of frames, no of frames usable) in first row, (dist from player 1 hand, dist from player 2 hand, player 1 hand id) in second row, (player 2 hand id, 0, 0) in third row and (0,0,0) in all other rows), hit info ((is contact with player 1 racket?, is contact with player 2 racket?, is contact with table?)), next 44 columns contain 3d feature points of player 1, next 44 columns contain 3d feature points of player 2, next columns contain ball position in 3d