Skip to content

Latest commit

 

History

History
182 lines (138 loc) · 6.4 KB

README.md

File metadata and controls

182 lines (138 loc) · 6.4 KB
Raw

Semantic Human Mesh Reconstruction with Textures

Xiaoyu Zhan · Jianxin Yang · Yuanqi Li · Jie Guo · Yanwen Guo · Wenping Wang

Project Page Arxiv Link


This repository contains the official Pytorch implementation for Semantic Human Mesh Reconstruction with Textures.

teaser

Overview

📍 Installation

1. Get Started

Start from creating a conda environment.

git clone https://github.com/ZhanxyR/SHERT.git
cd SHERT
conda create -n shert python=3.8
conda activate shert

2. Install Pytorch

Follow Pytorch.

We recommend to use Pytorch >= 2.0 (lowest tested on 1.13).

3. Install Open3D

Important

Please install a specific version of Open3D manually to avoid any problem.

(We will fix the bugs to adapt to higher versions later.)

pip install open3d==0.10.0

4. Install Other Dependencies

pip install -r requirements.txt

5. Build Pytorch3D

Follow Pytorch3D. We recommend to build from source code.

The version we used is v0.7.6, while the lower versions should also be applicable.

Note

If you have any trouble in building the package, you could just set the refine_iter to 1 in corresponding config.yaml to avoid using Pytorch3D.
(e.g. ./examples/demo_scan/config.yaml)

6. Download Required Data and Checkpoints

Selective to dowload data/smplx, data/face, examples/*, and save/ckpt from NJU-Box or Google-Drive.

Dowload SMPL-X v1.1 Models (Male, Female, Neutral) from SMPL-X and put them to data/models.

The completed structure should be like:

|-- SHERT
    |-- data
        |-- cameras
        |-- masks
        |-- smplx
        |-- face
        |-- models
            |-- smplx
                |-- SMPLX_*.npz
    |-- examples
        |-- demo_image_w_gt_smplx
        |-- demo_image
        |-- demo_scan
    |-- save
        |-- ckpt
            |-- inpaint.pth     # For mesh completion
            |-- refine.pth      # For mesh refinement
            |-- texture_local   # For texture inpainting
            |-- texture_global  # For texture repainting

🚀 Demo

  • The whole processes include two steps: reconstruction and texture inpainting.

⚡ Run quick_demo to test reconstruction in given resources. The results will be saved to ./examples/$subject$/results.

# Use ECON-pred mesh and fitted smplx.
python -m apps.quick_demo

# Use THuman scan and fitted smplx.
python -m apps.quick_demo -e scan

# Given only the image and predict all inputs with ECON.
python -m apps.quick_demo -e image

🖥️ For texture inpainting, we provide a client script and a server script that enables you to run diffusion model on remote server.

The client script will create a webui using Gradio, which can be accessed by http://localhost:7860.

Note

If you run the client in a new environment, some corresponding dependencies should be reinstalled.
When you first use inpainting, it will download the pretrained diffusion checkpoints from Huggingface.

# Run server
python -m apps.texture_rpc_server

# Run remote client
python -m apps.texture_client -i <server.ip>

# Run local client
python -m apps.texture_client -i localhost

🏵️ Acknowledgments

This work was supported by the National Natural Science Foundation of China (No. 62032011) and the Natural Science Foundation of Jiangsu Province (No. BK20211147).

There are also many powerful resources that greatly benefit our work:

🖋️ Citation

@inproceedings{zhan2024shert,
    title     = {Semantic Human Mesh Reconsturction with Textures},
    author    = {Zhan, Xiaoyu and Yang, Jianxin and Li, Yuanqi and Guo, Jie and Guo, Yanwen and Wang, Wenping},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2024},
}

📧 Contact

Zhan, Xiaoyu ([email protected]) and Yang, Jianxin ([email protected])