dataset_info.json

{
  "dataset_name": "ViLAP-Vietnamese",
  "description": "Vietnamese Landmark-Audio-Pose dataset for talking-face landmark generation. Collected from public YouTube Vietnamese news/interview videos. Each clip contains 03_face (cropped face images), 05_landmark (per-frame .npy), and 06_audio (audio.npy mel + audio.wav).",
  "version": "1.0.0",
  "split_seed": 42,
  "splits": {
    "train": {
      "clips": 3541,
      "speakers": 123
    },
    "val": {
      "clips": 338,
      "speakers": 105
    },
    "test": {
      "clips": 239,
      "speakers": 93
    }
  },
  "total": {
    "clips": 4118,
    "speakers": 123
  },
  "sources": {
    "vilap_dataset": 1630,
    "vilap_dataset_extra_10h": 2488
  },
  "clip_format": "speaker_id/speaker_id_cNNNN",
  "data_dirs": {
    "02_clips": "Raw .mp4 clips (25fps, 1280x720)",
    "03_face": "Cropped face images 128x128 .png, one per frame",
    "04_sketch": "Sketch images for reference (optional)",
    "05_landmark": "Per-frame .npy files with pose_landmarks (74 pts) and content_landmarks (57 pts), normalized to [0,1] in face crop",
    "06_audio": "audio.npy (mel spectrogram, shape [T,80]) and audio.wav (16kHz mono)"
  },
  "landmark_convention": {
    "coordinate_system": "Normalized to face crop [0.0, 1.0]",
    "format": "list of [landmark_id, x, y]",
    "pose_landmarks": "74 points — jaw + upper face structure",
    "content_landmarks": "57 points — lower face (jaw 0:17, outer lip 17:37, inner lip 37:57)"
  },
  "audio_convention": {
    "sample_rate": 16000,
    "mel_bands": 80,
    "hop_size": 200,
    "win_size": 800,
    "n_fft": 800,
    "frame_rate": 25
  }
}