FaceFormer Emo

Abstract

We propose FaceFormer Emo, a Transformer-based speech to 3D face mesh model that produces highly emotional expressions from the input audio.

This work extends FaceFormer in producing realistic and expressive outputs without a large loss in lip vertex accuracy, and introduces a novel lip vertex loss function which increases the weights of the vertices nearby the lips to increase accuracy and emotion dynamics.

Poster

3-min Video Introduction

[YouTube Link]

FaceFormer

PyTorch implementation for the paper:

FaceFormer: Speech-Driven 3D Facial Animation with Transformers, CVPR 2022.

Yingruo Fan, Zhaojiang Lin, Jun Saito, Wenping Wang, Taku Komura

[Paper] [Project Page]

Given the raw audio input and a neutral 3D face mesh, our proposed end-to-end Transformer-based architecture, FaceFormer, can autoregressively synthesize a sequence of realistic 3D facial motions with accurate lip movements.

Environment

• Ubuntu 18.04.1
• Python 3.7
• Pytorch 1.9.0

Dependencies

• Check the required python packages in requirements.txt.
• ffmpeg
• MPI-IS/mesh

Data

VOCASET

Request the VOCASET data from https://voca.is.tue.mpg.de/. Place the downloaded files data_verts.npy, raw_audio_fixed.pkl, templates.pkl and subj_seq_to_idx.pkl in the folder VOCASET. Download "FLAME_sample.ply" from voca and put it in VOCASET/templates.

BIWI

Request the BIWI dataset from Biwi 3D Audiovisual Corpus of Affective Communication. The dataset contains the following subfolders:

• 'faces' contains the binary (.vl) files for the tracked facial geometries.
• 'rigid_scans' contains the templates stored as .obj files.
• 'audio' contains audio signals stored as .wav files.

Place the folders 'faces' and 'rigid_scans' in BIWI and place the wav files in BIWI/wav.

Demo

Download the pretrained models from biwi.pth and vocaset.pth. Put the pretrained models under BIWI and VOCASET folders, respectively. Given the audio signal,

• to animate a mesh in BIWI topology, run:

  python demo.py --model_name biwi --wav_path "demo/wav/test.wav" --dataset BIWI --vertice_dim 70110  --feature_dim 128 --period 25 --fps 25 --train_subjects "F2 F3 F4 M3 M4 M5" --test_subjects "F1 F5 F6 F7 F8 M1 M2 M6" --condition M3 --subject M1

• to animate a mesh in FLAME topology, run:

  python demo.py --model_name vocaset --wav_path "demo/wav/test.wav" --dataset vocaset --vertice_dim 15069 --feature_dim 64 --period 30  --fps 30  --train_subjects "FaceTalk_170728_03272_TA FaceTalk_170904_00128_TA FaceTalk_170725_00137_TA FaceTalk_170915_00223_TA FaceTalk_170811_03274_TA FaceTalk_170913_03279_TA FaceTalk_170904_03276_TA FaceTalk_170912_03278_TA" --test_subjects "FaceTalk_170809_00138_TA FaceTalk_170731_00024_TA" --condition FaceTalk_170913_03279_TA --subject FaceTalk_170809_00138_TA

  This script will automatically generate the rendered videos in the demo/output folder. You can also put your own test audio file (.wav format) under the demo/wav folder and specify the argument --wav_path "demo/wav/test.wav" accordingly.

Training and Testing on VOCASET

Data Preparation

• Read the vertices/audio data and convert them to .npy/.wav files stored in vocaset/vertices_npy and vocaset/wav:

  cd VOCASET
  python process_voca_data.py

Training and Testing

(WARNING: The snippet below might exhaust all memory on GPU, use 5 train_subjects for safe measure)

• To train the model on VOCASET and obtain the results on the testing set, run:

  python main.py --dataset vocaset --vertice_dim 15069 --feature_dim 64 --period 30 --train_subjects "FaceTalk_170728_03272_TA FaceTalk_170904_00128_TA FaceTalk_170725_00137_TA FaceTalk_170915_00223_TA FaceTalk_170811_03274_TA FaceTalk_170913_03279_TA FaceTalk_170904_03276_TA FaceTalk_170912_03278_TA" --val_subjects "FaceTalk_170811_03275_TA FaceTalk_170908_03277_TA" --test_subjects "FaceTalk_170809_00138_TA FaceTalk_170731_00024_TA"

  The results and the trained models will be saved to vocaset/result and vocaset/save.

Visualization

• To visualize the results, run:

  python render.py --dataset vocaset --vertice_dim 15069 --fps 30

  You can find the outputs in the vocaset/output folder.

Acknowledgement

We gratefully acknowledge ETHZ-CVL for providing the B3D(AC)2 database and MPI-IS for releasing the VOCASET dataset. The implementation of wav2vec2 is built upon huggingface-transformers, and the temporal bias is modified from ALiBi. We use MPI-IS/mesh for mesh processing and VOCA/rendering for rendering. We thank the authors for their excellent works. Any third-party packages are owned by their respective authors and must be used under their respective licenses.