TalkingAvatar Dataset for Paper THGS

This document provides detailed steps on how to access and use data hosted on Google Drive and Baidu Netdisk.

Google Drive

1. Visit the link: Google Drive

Baidu Netdisk

1. Visit the link: Baidu Netdisk
2. Enter the password: THGS

Fah Dataset Structure Overview

The Fah dataset is organized into multiple directories, each corresponding to a 10-second video segment. Below is a detailed breakdown of the dataset's structure and contents.

Data Structure Example

Fah/
├── Fah_output000/ (data folder corresponding to a 10s video)
│   ├── Fah_output000.mp4
│   ├── aud.wav
│   ├── aud_eo.npy (audio feature)
│   ├── deca/
│   ├── deca_vis/
│   ├── face_mask/
│   ├── fan/
│   ├── image (ground truth)
│   ├── keypoints.npy (2d keypoints)
│   ├── ours_exp (estimated SMPL-X params)
│   ├── parsing (foreground mask)
│   ├── pixie/
│   └── pixie_vis/
├── Fah_output001/ (similar data folder structure)
├── Fah_output002/ (similar data folder structure)
├── Fah_output003/ (similar data folder structure)
...
├── optimized_camera_params.npz (camera parameters)
└── poses_optimized.npz (refined body parameters through Joint Optimization Strategy, Stage 1)

Example for using the data

Once the data is downloaded, it can be decompressed in Linux using the following command:

tar -xzvf Fah.tar.gz 

videoDataset.py provide an api to parse the data. Change root_dir to your root directory, subject to 'Fah' for example, opt settings can be found in train.yaml.

Important Notes

Ensure compliance with all relevant usage and distribution policies. If there are any problems, please contact [[email protected]] or directly raise an issue.