README.md

---
language:
- vi
license: cc-by-nc-4.0
task_categories:
- video-classification
- image-to-video
tags:
- talking-face
- landmark
- lip-sync
- vietnamese
- audio-visual
pretty_name: 'VAVD: Vietnamese Visual-Audio Dataset'
size_categories:
- 1K<n<10K
---

# Vietnamese Visual-Audio Dataset

Vietnamese audio-visual landmark dataset designed for talking-face landmark generation. The dataset is curated to support efficient fine-tuning of the pretrained [IP-LAP](https://github.com/Weizhi-Zhong/IP_LAP) model on Vietnamese speakers.

> **Note:** This dataset does not directly host video files. Instead, [`video_sources.csv`](video_sources.csv) provides the 123 public YouTube source URLs. The full dataset can be automatically downloaded and processed using [`download_dataset.py`](download_dataset.py).

---

## Dataset Summary

| Split | Clips | Speakers |
|-------|------:|--------:|
| Train | 3,541 | ~118 |
| Val | 338 | ~66 |
| Test | 239 | ~52 |
| **Total** | **4,118** | **123** |

The dataset is constructed from two sources:

- `vilap_dataset`: 1,630 clips from the main training set
- `vilap_dataset_extra_10h`: 2,488 clips from the extended crawl

The videos are collected from public Vietnamese YouTube sources, including news programs, interviews, and talk shows. Each sample is a short single-speaker clip of approximately 5 seconds, processed at 25 FPS.

---

## Quick Start

```bash
# 1. Clone the dataset repository from Hugging Face
git clone https://huggingface.co/datasets/btlam2002/VAVD
cd VAVD

# 2. Install the required dependencies
pip install yt-dlp mediapipe opencv-python librosa scipy numpy

# 3. Download and process the full dataset
# This automatically downloads videos, cuts clips, extracts faces, landmarks, and audio features
python download_dataset.py --output ./vilap_data

# Or download only the test split for a quick trial (~239 clips)
python download_dataset.py --output ./vilap_data --splits test

# Or process only specific videos
python download_dataset.py --output ./vilap_data --video_ids 3p7dFrIx5bk xwsPD6xiPbI
```

The `download_dataset.py` pipeline automatically performs the following steps:

1. **Download videos** from YouTube using `yt-dlp` with a maximum resolution of 720p.
2. **Cut clips** into 5-second segments using `ffmpeg`.
3. **Extract face crops** at 128×128 pixels using MediaPipe Face Detection.
4. **Extract landmarks** using MediaPipe Face Mesh, including 74 pose landmarks and 57 content landmarks.
5. **Extract audio features**, including 16 kHz WAV files and Mel-spectrogram `.npy` files following the IP-LAP convention.

---

## Video Sources

All video URLs are listed in [`video_sources.csv`](video_sources.csv).

| Column | Description |
|--------|-------------|
| `video_id` | YouTube video ID, also used as the speaker folder name |
| `youtube_url` | Full YouTube URL |
| `title` | Video title |
| `channel` | YouTube channel name |
| `source` | Dataset source, either `vilap_dataset` or `vilap_dataset_extra_10h` |
| `n_train_clips` | Number of clips in the training split |
| `n_val_clips` | Number of clips in the validation split |
| `n_test_clips` | Number of clips in the test split |
| `n_total_clips` | Total number of clips from the video |

Example:

```csv
video_id,youtube_url,title,channel,source,n_train_clips,n_val_clips,n_test_clips,n_total_clips
3p7dFrIx5bk,https://www.youtube.com/watch?v=3p7dFrIx5bk,Chuyên gia phân tích...,VTV24,vilap_dataset,68,1,1,70
```

---

## Directory Structure

```text
vilap_dataset/
├── 02_clips/
│   └── <speaker_id>/
│       └── <speaker_id>_cNNNN.mp4       # raw clip, 25 FPS, 1280×720
├── 03_face/
│   └── <speaker_id>/<clip_id>/
│       └── {frame_id}.png               # 128×128 face crop
├── 04_sketch/
│   └── <speaker_id>/<clip_id>/
│       └── {frame_id}.png               # optional sketch image
├── 05_landmark/
│   └── <speaker_id>/<clip_id>/
│       └── {frame_id}.npy               # per-frame landmark dictionary
└── 06_audio/
    └── <speaker_id>/<clip_id>/
        ├── audio.npy                    # Mel-spectrogram, shape [T, 80]
        └── audio.wav                    # 16 kHz mono WAV file
```

---

## Landmark Format

Each `.npy` landmark file contains a Python dictionary and can be loaded with `allow_pickle=True`.

```python
d = np.load("05_landmark/spk/spk_c0000/0.npy", allow_pickle=True).item()
# d.keys() → ['pose_landmarks', 'content_landmarks']
```

| Field | Points | Description |
|-------|-------:|-------------|
| `pose_landmarks` | 74 | Jaw and upper-face structural landmarks |
| `content_landmarks` | 57 | Lip and lower-face landmarks, including jaw `0:17`, outer lip `17:37`, and inner lip `37:57` |

Each landmark is stored as a list of `[landmark_id, x, y]` triplets, where `x` and `y` are normalized to the face crop range `[0.0, 1.0]`.

```python
# Example: read one frame
import numpy as np

d = np.load("frame_0.npy", allow_pickle=True).item()

content = d["content_landmarks"]   # list of [id, x, y], 57 points
pose = d["pose_landmarks"]         # list of [id, x, y], 74 points

# Sort landmarks according to the IP-LAP ori_sequence_idx before feeding them into the model.
```

---

## Audio Format

| Property | Value |
|----------|-------|
| Sample rate | 16,000 Hz |
| Channels | Mono |
| Mel bands | 80 |
| FFT size | 800 |
| Hop size | 200 samples |
| Frame rate | 25 FPS, aligned with video |

```python
# Load Mel-spectrogram
mel = np.load("audio.npy")   # shape: [T_mel_frames, 80]

# Extract a 16-frame Mel window for frame i following the IP-LAP convention
frame_idx = 50
audio_offset = -2           # -4 recommended offset for Vietnamese speech

start = int(80.0 * ((frame_idx + audio_offset) / 25.0))
window = mel[start : start + 16, :]   # shape: [16, 80]
```

---

## File Lists

The `filelists/` directory provides pre-computed train, validation, and test splits.

```text
filelists/
├── train.txt   # 3,541 clips, 86.0%
├── val.txt     # 338 clips, 8.2%
└── test.txt    # 239 clips, 5.8%
```

Each line stores a clip key in the format `<speaker_id>/<clip_id>`.

```text
3p7dFrIx5bk/3p7dFrIx5bk_c0000
3p7dFrIx5bk/3p7dFrIx5bk_c0001
...
```

The following example shows how to resolve the actual data paths:

```python
from pathlib import Path

DATA_BASE = Path("path/to/vilap_dataset")

clip_key = "3p7dFrIx5bk/3p7dFrIx5bk_c0017"
dirname, vidname = clip_key.split("/")

face_dir = DATA_BASE / "03_face" / dirname / vidname
lm_dir = DATA_BASE / "05_landmark" / dirname / vidname
audio_npy = DATA_BASE / "06_audio" / dirname / vidname / "audio.npy"
audio_wav = DATA_BASE / "06_audio" / dirname / vidname / "audio.wav"
```

---

## Loading a Full Clip

```python
import numpy as np
import torch
from pathlib import Path

DATA_BASE = Path("vilap_dataset")

ORI_SEQUENCE_IDX = [
    162, 127, 234, 93, 132, 58, 172, 136, 150, 149, 176, 148, 152,
    377, 400, 378, 379, 365, 397, 288, 361, 323, 454, 356, 389,
    70, 63, 105, 66, 107, 55, 65, 52, 53, 46,
    336, 296, 334, 293, 300, 276, 283, 282, 295, 285,
    168, 6, 197, 195, 5,
    48, 115, 220, 45, 4, 275, 440, 344, 278,
    33, 246, 161, 160, 159, 158, 157, 173, 133, 155, 154, 153, 145, 144, 163, 7,
    362, 398, 384, 385, 386, 387, 388, 466, 263, 249, 390, 373, 374, 380, 381, 382,
    61, 185, 40, 39, 37, 0, 267, 269, 270, 409, 291, 375, 321, 405, 314, 17, 84, 181, 91, 146,
    78, 191, 80, 81, 82, 13, 312, 311, 310, 415, 308, 324, 318, 402, 317, 14, 87, 178, 88, 95,
]


def load_clip(clip_key, data_base):
    dirname, vidname = clip_key.split("/")
    lm_dir = data_base / "05_landmark" / dirname / vidname

    npys = sorted(lm_dir.glob("*.npy"), key=lambda p: int(p.stem))
    frame_ids = [int(p.stem) for p in npys]

    pose_list, content_list = [], []

    for fid in frame_ids:
        d = np.load(lm_dir / f"{fid}.npy", allow_pickle=True).item()

        for key, pts, n in [
            ("pose_landmarks", pose_list, 74),
            ("content_landmarks", content_list, 57),
        ]:
            lm = sorted(d[key], key=lambda t: ORI_SEQUENCE_IDX.index(int(t[0])))

            arr = torch.zeros(2, n)
            arr[0] = torch.tensor([float(t[1]) for t in lm])
            arr[1] = torch.tensor([float(t[2]) for t in lm])

            pts.append(arr)

    all_pose = torch.stack(pose_list)        # shape: [N, 2, 74]
    all_content = torch.stack(content_list)  # shape: [N, 2, 57]

    mel = np.load(data_base / "06_audio" / dirname / vidname / "audio.npy")

    return all_pose, all_content, mel, frame_ids


pose, content, mel, fids = load_clip(
    "3p7dFrIx5bk/3p7dFrIx5bk_c0017",
    DATA_BASE,
)

print(
    f"Frames: {len(fids)}, "
    f"pose: {pose.shape}, "
    f"content: {content.shape}, "
    f"mel: {mel.shape}"
)
```

---

## Processing New Videos

To generate the dataset from raw videos, use the pipeline provided in [`scripts/process_vilap_data_pipeline.py`](../../scripts/process_vilap_data_pipeline.py).

```text
Step 1: Download or prepare raw .mp4 videos  →  02_clips/
Step 2: Detect and crop faces                →  03_face/
Step 3: Generate sketches, optional           →  04_sketch/
Step 4: Extract MediaPipe landmarks           →  05_landmark/
Step 5: Extract and resample audio            →  06_audio/
```

Requirements:

- Python 3.10.2
- `mediapipe`
- `librosa`
- `opencv-python`
- `ffmpeg`

For Mel-spectrogram extraction details consistent with IP-LAP, see [`scripts/preprocess_audio.py`](../../third_party/IP_LAP/preprocess/preprocess_audio.py).

---
<!-- 
## Model

The ViLAP model is a LoRA-based fine-tuned version of [IP-LAP](https://github.com/Weizhi-Zhong/IP_LAP) trained on the proposed Vietnamese dataset. Model checkpoints are released separately.

| Model | LMD ↓ | Mouth-MAE ↓ | Pearson-r ↑ |
|-------|------:|------------:|------------:|
| IP-LAP pretrained | 0.01386 | 0.01737 | 0.462 |
| ** LoRA Talk** | **0.01139** | **0.01357** | **0.522** |
 -->
---
<!-- 
## Citation

```bibtex
@misc{vilap2025,
  title = {ViLAP: Vietnamese Landmark-Audio-Pose Dataset for Talking Face Generation},
  year = {2025},
  note = {Hugging Face dataset},
}
```
 -->
---

## License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International License, CC BY-NC 4.0.

The original videos are sourced from publicly available YouTube content. The dataset is intended for research and non-commercial use only.