This is an official repository for the paper

DAD-3DHeads: A Large-scale Dense, Accurate and Diverse Dataset for 3D Head Alignment from a Single Image
Tetiana Martyniuk, Orest Kupyn, Yana Kurlyak, Igor Krashenyi, Jiři Matas, Viktoriia Sharmanska
CVPR 2022

The code uses Python 3.8.

conda create --name DAD-3DHeads python=3.6
conda activate DAD-3DHeads

git clone https://github.com/PinataFarms/DAD-3DHeads.git
cd DAD-3DHeads

pip install -r requirements.txt

First, you need to download the DAD-3DHeads dataset and extract it to the dataset/DAD-3DHeadsDataset directory. The dataset is available upon request. Please fill in this form to get access to it.

The dataset directory structure should be the following:

./dataset
--DAD-3DHeadsDataset
----train
------images
--------<ID>.png
------annotations
--------<ID>.json
------train.json
----val
------images/<ID>.png
------annotations/<ID>.json
------val.json
----test
------images/<ID>.png
------test.json

Annotations <ID>.json file structure:

--vertices
--model_view_matrix
--projection_matrix

Metadata [train|val|test].json file structure:

--item_id
--annotation_path
--img_path
--bbox #[x, y, w, h] format
----0
----1
----2
----3
--attributes
----quality #[hq, lq]
----gender #[female, male, undefined]
----expression #[true, false]
----age #[child, young, middle_aged, senior]
----occlusions #[true, false]
----pose #[front, sided, atypical]
----standard light #[true, false]

The training code uses hydra. To change the training setup, add a new or edit the existing .yaml file in the model_training/config folder.

python visualize.py <subset> <id>

Pick subset from the train, val, test options, and the corresponding item_id (without file extension).

python train.py

First row (from left to right): input image, 68 2D face landmarks visualization, 191 2D face landmarks visualization, 445 2D face landmarks visualization.
Second row (from left to right): face mesh visualization, head mesh visualization, head pose visualization, 3D head mesh.

python demo.py <path/to/input/image.png> <path/to/output/folder> <type_of_output>

# Visualize 68 2D face landmarks
python demo.py images/demo_heads/1.jpeg outputs 68_landmarks

# Visualize 191 2D face landmarks
python demo.py images/demo_heads/1.jpeg outputs 191_landmarks

# Visualize 445 2D face landmarks
python demo.py images/demo_heads/1.jpeg outputs 445_landmarks

# Visualize face mesh
python demo.py images/demo_heads/1.jpeg outputs face_mesh

# Visualize head mesh
python demo.py images/demo_heads/1.jpeg outputs head_mesh

# Visualize head pose
python demo.py images/demo_heads/1.jpeg outputs pose

# Get 3D mesh .obj file
python demo.py images/demo_heads/1.jpeg outputs 3d_mesh

# Get flame parameters .json file
python demo.py images/demo_heads/1.jpeg outputs flame_params

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.

By using this code, you acknowledge that you have read the license terms, understand them, and agree to be bound by them. If you do not agree with these terms and conditions, you must not use the code.

If you use the DAD-3DHeads Dataset and/or this code - implicitly or explicitly - for your research projects, please cite the following paper:

@inproceedings{dad3dheads,
    title={DAD-3DHeads: A Large-scale Dense, Accurate and Diverse Dataset for 3D Head Alignment from a Single Image},
    author={Martyniuk, Tetiana and Kupyn, Orest and Kurlyak, Yana and Krashenyi, Igor and Matas, Ji\v{r}i and Sharmanska, Viktoriia},
    booktitle = {Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR)},
    year={2022}
}