[ECCV 2020] Reimplementation of 3DDFAv2, including face mesh, head pose, landmarks, and more.

Overview

Stable Head Pose Estimation and Landmark Regression via 3D Dense Face Reconstruction

FaceReconstructionDemo

Language grade: Python License ECCV

Reimplementation of (ECCV 2020) Towards Fast, Accurate and Stable 3D Dense Face Alignment via Tensorflow Lite framework, face mesh, head pose, landmarks, and more.

  • CPU real-time face deteciton, alignment, and reconstruction pipeline.
  • Lightweight render library, 5x faster (3ms vs 15ms) than the Sim3DR tools.
  • Camera matrix and dense/sparse landmarks prediction via a single network.
  • Generate facial parameters for robust head pose and expression estimation.

Setup

Basic Requirements

  • Python 3.6+
  • pip3 install -r requirements.txt

Render for Dense Face

  • GCC 6.0+
  • bash build_render.sh
  • (Cautious) For Windows user, please refer to this tutorial for more details.

3D Facial Landmarks

In this project, we perform dense face reconstruction by 3DMM parameters regression. The regression target is simplified as camera matrix (C, shape of 3x4), appearance parameters (S, shape of 1x40), and expression variables (E, shape of 1x10), with 62 dimensions in total.

The sparse or dense facial landmarks can be estimated by applying these parameters to a predefined 3D model, such as BFM. More specifically, the following formula describes how to generate a face through parameters:

Generate Face

where U and W are from pre-defined face model. Combine them linearly with parameters to generate sparse or dense faces. Finally, we need to integrate the posture information into the result:

With matrix

where R (shape of 3x3) and T (shape of 3x1) denote rotation and translation matrices, respectively, which are fractured from the camera matrix C.

Sparse

sparse demo

Since we have reconstructed the entire face, the 3D face alignment can be achieved by selecting the landmarks at the corresponding positions. See [TPAMI 2017] Face alignment in full pose range: A 3d total solution for more details.

Comparing with the method of first detecting 2D landmarks and then performing depth estimation, directly fitting 3DMM to solve 3D face alignment can not only obtain more accurate results in larger pose scenes, but also has obvious advantages in speed.

We provide a demonstration script that can generate 68 landmarks based on the reconstructed face. Run the following command to view the real-time 3D face alignment results:

python3 demo_video.py -m sparse -f <your-video-path>

Dense

dense demo

Currently, our method supports up to 38,365 landmarks. We draw landmarks every 6 indexes for a better illustration. Run the demonstrate script in dense mode for real-time dense facial landmark localization:

python3 demo_video.py -m dense -f <your-video-path>

Face Reconstruction

Our network is multi-task since it can directly regress 3DMM params from a single face image for reconstruction, as well as estimate the head pose via R and T prediction. During training, the predicted R can supervise the params regression branch to generate refined face mesh. Meanwhile, the landmarks calculated via the params are provided as the labeled data to the pose estimation branch for training, through the cv2.solvePnP tool.

Theoretically speaking, the head pose estimation task in our model is weakly supervised, since only a few labeled data is required to activate the training process. In detail, the loss function at the initial stage can be described as follows:

Init Pose Loss

After the initialization process, the ground truth can be replaced by the prediction results of other branches. Therefore, the loss function will be transformed to the following equation in the later stage of the training process:

Pose Loss

In general, fitting the 3D model during the training process dynamically can avoid the inaccurate head pose estimation results caused by the coarse predefined model.

Head Pose

pose demo

Traditional head pose estimation approaches, such as Appearance Template Models, Detector Arrays, and Mainfold Embedding have been extensively studied. However, methods based on deep learning improved the prediction accuracy to meet actual needs, until recent years.

Given a set of predefined 3D facial landmarks and the corresponding 2D image projections, the SolvePnP tool can be utilized to calculate the rotation matrix. However, the adopted mean 3D human face model usually introduces intrinsic error during the fitting process. Meanwhile, the additional landmarks extraction component is also kind of cumbersome.

Therefore, we designed a network branch for directly regress 6DoF parameters from the face image. The predictions include the 3DoF rotation matrix R (Pitch, Yaw, Roll), and the 3DoF translation matrix T (x, y, z), Compared with the landmark-based method, directly regression the camera matrix is more robust and stable, as well as significantly reduce the network training cost. Run the demonstrate script in pose mode to view the real-time head pose estimation results:

python3 demo_video.py -m pose -f <your-video-path>

Expression

Expression

Coarse expression estimation can be achieved by combining the predefined expressions in BFM linearly. In our model, regression the E is one of the tasks of the params prediction branch. Obviously, the accuracy of the linear combination is positively related to the dimension.

Clipping parameters can accelerate the training process, however, it can also reduce reconstruction accuracy, especially details such as eye and mouth. More specifically, E has a greater impact on face details than S when emotion is involved. Therefore, we choose 10-dimension for a tradeoff between the speed and the accuracy, the training data can be found at here for refinement.

In addition, we provide a simple facial expression rendering script. Run the following command for illustration:

python3 demo_image.py <your-image-path>

Mesh

mesh demo

According to the predefined BFM and the predicted 3DMM parameters, the dense 3D facial landmarks can be easily calculated. On this basis, through the index mapping between the morphable triangle vertices and the dense landmarks defined in BFM, the renderer can plot these geometries with depth infomation for mesh preview. Run the demonstrate script in mesh mode for real-time face reconstruction:

python3 demo_video.py -m mesh -f <your-video-path>

Benchmark

Our network can directly output the camera matrix and sparse or dense landmarks. Compared with the model in the original paper with the same backbone, the additional parameters yield via the pose regression branch does not significantly affect the inference speed, which means it can still be CPU real-time.

Scheme THREAD=1 THREAD=2 THREAD=4
Inference 7.79ms 6.88ms 5.83ms

In addition, since most of the operations are wrapped in the model, the time consumption of pre-processing and post-processing are significantly reduced. Meanwhile, the optimized lightweight renderer is 5x faster (3ms vs 15ms) than the Sim3DR tools. These measures decline the latency of the entire pipeline.

Stage Preprocess Inference Postprocess Render
Each face cost 0.23ms 7.79ms 0.39ms 3.92ms

Run the following command for speed benchmark:

python3 video_speed_benchmark.py <your-video-path>

Citation

@inproceedings{guo2020towards,
    title={Towards Fast, Accurate and Stable 3D Dense Face Alignment},
    author={Guo, Jianzhu and Zhu, Xiangyu and Yang, Yang and Yang, Fan and Lei, Zhen and Li, Stan Z},
    booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
    year={2020}
}
Comments
  • __init__() got an unexpected keyword argument 'num_threads'

    __init__() got an unexpected keyword argument 'num_threads'

    While running "python3 demo_video.py -m mesh -f I receive the following ERROR: File "demo_video.py", line 53, in main(args) File "demo_video.py", line 16, in main fa = service.DenseFaceReconstruction("weights/dense_face.tflite") File "/home/happy/PycharmProjects/Dense-Head-Pose-Estimation/service/TFLiteFaceAlignment.py", line 11, in init num_threads=num_threads) TypeError: init() got an unexpected keyword argument 'num_threads'

    opened by berylyellow 2
  • asset/render.so not found

    asset/render.so not found

    While running "python3 demo_video.py -m mesh -f I receive the following ERROR OSError: asset/render.so: cannot open shared object file: No such file or directory I can successfully run the other three modes (sparse, dense, pose)

    opened by fisakhan 2
  • How can i get Depth

    How can i get Depth

    hi, I want to get information about the depth of the face, how do I calculate it? i don't need rendering but only need depth map. like https://github.kakaocorp.com/data-sci/3DDFA_V2_tf2/blob/master/utils/depth.py thank you in advance!

    opened by joongkyunKim 1
  • there is no asset/render.so

    there is no asset/render.so

    Ignore above cudart dlerror if you do not have a GPU set up on your machine. Traceback (most recent call last): File "demo_video.py", line 53, in main(args) File "demo_video.py", line 18, in main color = service.TrianglesMeshRender("asset/render.so", File "/home/xuhao/workspace/Dense-Head-Pose-Estimation-main/service/CtypesMeshRender.py", line 17, in init self._clibs = ctypes.CDLL(clibs)

    opened by hitxuhao 1
  • asset/render.so: cannot open shared object file: No such file or director

    asset/render.so: cannot open shared object file: No such file or director

    It worked well with: $ python3 demo_video.py -m pose -f ./re.mp4 $python3 demo_video.py -m sparse -f ./re.mp4 $ python3 demo_video.py -m dense -f ./re.mp4

    but with mesh not as you see here:

    $python3 demo_video.py -m mesh -f ./re.mp4 2022-12-05 17:03:03.713207: I tensorflow/stream_executor/platform/default/dso_loader.cc:48] Successfully opened dynamic library libcudart.so.10.1 Traceback (most recent call last): File "demo_video.py", line 55, in main(args) File "demo_video.py", line 21, in main color = service.TrianglesMeshRender("asset/render.so", "asset/triangles.npy") File "/home/redhwan/2/HPE/tensorflow/Dense-Head-Pose-Estimation-main/service/CtypesMeshRender.py", line 17, in init self._clibs = ctypes.CDLL(clibs) File "/usr/lib/python3.8/ctypes/init.py", line 373, in init self._handle = _dlopen(self._name, mode) OSError: asset/render.so: cannot open shared object file: No such file or directory

    opened by Algabri 0
  • Memory usage

    Memory usage

    hi @1996scarlet

    Is it expected that the model uses around 18GB of video RAM? Or does it maximises the RAM available to help with speed? (I got a 24GB ram GPU, and usage goes up to 22.2GB while inferencing) Thanks for this amazing MIT license repo btw.

    opened by Tetsujinfr 0
Releases(v1.0)
  • v1.0(Feb 1, 2021)

    Stable Head Pose Estimation and Landmark Regression via 3D Dense Face Reconstruction

    • CPU real-time face deteciton, alignment, and reconstruction pipeline.
    • Lightweight render library, 5x faster (3ms vs 15ms) than the Sim3DR tools.
    • Camera matrix and dense/sparse landmarks prediction via a single network.
    • Generate facial parameters for robust head pose and expression estimation.
    Source code(tar.gz)
    Source code(zip)
Owner
Remilia Scarlet
Remilia Scarlet
Dynamic wallpaper generator.

Wiki • About • Installation About This project is a dynamic wallpaper changer. It waits untill you turn on the music, downloads album cover if it's po

3 Sep 18, 2021
Real-Time Multi-Contact Model Predictive Control via ADMM

Here, you can find the code for the paper 'Real-Time Multi-Contact Model Predictive Control via ADMM'. Code is currently being cleared up and optimize

17 Dec 28, 2022
Ranking Models in Unlabeled New Environments (iccv21)

Ranking Models in Unlabeled New Environments Prerequisites This code uses the following libraries Python 3.7 NumPy PyTorch 1.7.0 + torchivision 0.8.1

14 Dec 17, 2021
Permute Me Softly: Learning Soft Permutations for Graph Representations

Permute Me Softly: Learning Soft Permutations for Graph Representations

Giannis Nikolentzos 7 Jul 10, 2022
Unofficial implementation of MLP-Mixer: An all-MLP Architecture for Vision

MLP-Mixer: An all-MLP Architecture for Vision This repo contains PyTorch implementation of MLP-Mixer: An all-MLP Architecture for Vision. Usage : impo

Rishikesh (ऋषिकेश) 175 Dec 23, 2022
[NeurIPS 2021] Better Safe Than Sorry: Preventing Delusive Adversaries with Adversarial Training

Better Safe Than Sorry: Preventing Delusive Adversaries with Adversarial Training Code for NeurIPS 2021 paper "Better Safe Than Sorry: Preventing Delu

Lue Tao 29 Sep 20, 2022
RoIAlign & crop_and_resize for PyTorch

RoIAlign for PyTorch This is a PyTorch version of RoIAlign. This implementation is based on crop_and_resize and supports both forward and backward on

Long Chen 530 Jan 07, 2023
Code For TDEER: An Efficient Translating Decoding Schema for Joint Extraction of Entities and Relations (EMNLP2021)

TDEER (WIP) Code For TDEER: An Efficient Translating Decoding Schema for Joint Extraction of Entities and Relations (EMNLP2021) Overview TDEER is an e

Alipay 6 Dec 17, 2022
Official code repository for A Simple Long-Tailed Rocognition Baseline via Vision-Language Model.

BALLAD This is the official code repository for A Simple Long-Tailed Rocognition Baseline via Vision-Language Model. Requirements Python3 Pytorch(1.7.

peng gao 42 Nov 26, 2022
An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise Weight Sharing) by Sensetime Research.

An integration of several popular automatic augmentation methods, including OHL (Online Hyper-Parameter Learning for Auto-Augmentation Strategy) and AWS (Improving Auto Augment via Augmentation Wise

45 Dec 08, 2022
CTF challenges and write-ups for MicroCTF 2021.

MicroCTF 2021 Qualifications About This repository contains CTF challenges and official write-ups for MicroCTF 2021 Qualifications. License Distribute

Shellmates 12 Dec 27, 2022
Latex code for making neural networks diagrams

PlotNeuralNet Latex code for drawing neural networks for reports and presentation. Have a look into examples to see how they are made. Additionally, l

Haris Iqbal 18.6k Jan 01, 2023
Audio2Face - Audio To Face With Python

Audio2Face Discription We create a project that transforms audio to blendshape w

FACEGOOD 724 Dec 26, 2022
Official pytorch implementation of the paper: "SinGAN: Learning a Generative Model from a Single Natural Image"

SinGAN Project | Arxiv | CVF | Supplementary materials | Talk (ICCV`19) Official pytorch implementation of the paper: "SinGAN: Learning a Generative M

Tamar Rott Shaham 3.2k Dec 25, 2022
Python library containing BART query generation and BERT-based Siamese models for neural retrieval.

Neural Retrieval Embedding-based Zero-shot Retrieval through Query Generation leverages query synthesis over large corpuses of unlabeled text (such as

Amazon Web Services - Labs 35 Apr 14, 2022
Code for the paper titled "Prabhupadavani: A Code-mixed Speech Translation Data for 25 languages"

Prabhupadavani: A Code-mixed Speech Translation Data for 25 languages Code for the paper titled "Prabhupadavani: A Code-mixed Speech Translation Data

Ayush Daksh 12 Dec 01, 2022
Tiny-NewsRec: Efficient and Effective PLM-based News Recommendation

Tiny-NewsRec The source codes for our paper "Tiny-NewsRec: Efficient and Effective PLM-based News Recommendation". Requirements PyTorch == 1.6.0 Tensor

Yang Yu 3 Dec 07, 2022
Reading list for research topics in Masked Image Modeling

awesome-MIM Reading list for research topics in Masked Image Modeling(MIM). We list the most popular methods for MIM, if I missed something, please su

ligang 231 Dec 07, 2022
Winning solution of the Indoor Location & Navigation Kaggle competition

This repository contains the code to generate the winning solution of the Kaggle competition on indoor location and navigation organized by Microsoft

Tom Van de Wiele 62 Dec 28, 2022
(CVPR2021) ClassSR: A General Framework to Accelerate Super-Resolution Networks by Data Characteristic

ClassSR (CVPR2021) ClassSR: A General Framework to Accelerate Super-Resolution Networks by Data Characteristic Paper Authors: Xiangtao Kong, Hengyuan

Xiangtao Kong 308 Jan 05, 2023