search

xubaixinxbx / 3dheads

[ICCV 2023] Code for "Deformable Model-Driven Neural Rendering for High-Fidelity 3D Reconstruction of Human Heads Under Low-View Settings"
Other
39 stars 2 forks source link

readme

Deformable Model-Driven Neural Rendering for High-Fidelity 3D Reconstruction of Human Heads Under Low-View Settings

Code for "Deformable Model-Driven Neural Rendering for High-Fidelity 3D Reconstruction of Human Heads Under Low-View Settings"

Baixin Xu, Jiarui Zhang, Kwan-Yee Lin, Chen Qian, Ying He.

Project | Paper

Reconstructing 3D human heads in low-view settings presents technical challenges, mainly due to the pronounced risk of overfitting with limited views and high-frequency signals. To address this, we propose geometry decomposition and adopt a two-stage, coarse-to-fine training strategy, allowing for progressively capturing high-frequency geometric details. We represent 3D human heads using the zero level-set of a combined signed distance field, comprising a smooth template, a non-rigid deformation, and a high-frequency displacement field. The template captures features that are independent of both identity and expression and is co-trained with the deformation network across multiple individuals with sparse and randomly selected views. The displacement field, capturing individual-specific details, undergoes separate training for each person. Our network training does not require 3D supervision or object masks. Experimental results demonstrate the effectiveness and robustness of our geometry decomposition and two-stage training strategy. Our method outperforms existing neural rendering approaches in terms of reconstruction accuracy and novel view synthesis under low-view settings. Moreover, the pre-trained template serves a good initialization for our model when encountering unseen individuals.

Video

https://user-images.githubusercontent.com/113180125/226843465-432415fb-6ee3-44eb-930f-8d8a185f8a6d.mp4

Note: We upload the compressed video for convenience, please download and see the original video.

TODO

Our method is also evaluated on the Facescape dataset, which is processed as in NeuFace.

Setup

You can create an anaconda environment by referring to VolSDF.

Data

We use the data provided by Single Image Portrait Relighting via Explicit Multiple Reflectance Channel Modeling. Specifically, we select 30 distinct individuals and conduct joint training during stage 1.

Other datasets like Facescape please refer to the project page.

Training

Training template human head and deformation correspondence in stage 1:

bash confs/face_st1.sh

Training displacement field in stage 2 based on stage 1:

bash confs/face_st2.sh

Citation

If you find our work useful, please kindly cite as:

@InProceedings{Xu_2023_ICCV,
    author    = {Xu, Baixin and Zhang, Jiarui and Lin, Kwan-Yee and Qian, Chen and He, Ying},
    title     = {Deformable Model-Driven Neural Rendering for High-Fidelity 3D Reconstruction of Human Heads Under Low-View Settings},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
    month     = {October},
    year      = {2023},
    pages     = {17924-17934}
}

Acknowledgement