This repository is a reference implementation for NeuMan. NeuMan reconstructs both the background scene and an animatable human from a single video using neural radiance fields.
Novel view and novel pose synthesis
[Bike] [Citron] [Parking lot] [Jogging] [Lab] [Seattle]
Compositional Synthesis
To create the environment using Conda:
conda env create -f environment.ymlAlternately, you can create the environment by executing:
conda create -n neuman_env python=3.7 -y;
conda activate neuman_env;
conda install pytorch==1.8.0 torchvision==0.9.0 cudatoolkit=10.2 -c pytorch;
# For RTX 30 series GPU with CUDA version 11.x, please use:
# conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge
conda install -c fvcore -c iopath -c conda-forge fvcore iopath;
conda install -c bottler nvidiacub;
conda install pytorch3d -c pytorch3d;
conda install -c conda-forge igl;
pip install opencv-python joblib open3d imageio tensorboardX chumpy lpips scikit-image ipython matplotlib;Notice that pytorch3d requires a specific version of pytorch, in our case pytorch=1.8.0.
Activate the environment:
conda activate neuman_env-
Download SMPL weights:
- Registration is required to download the UV map(Download UV map in OBJ format) from SMPL.
- Download neutral SMPL weights(SMPLIFY_CODE_V2.ZIP) from SMPLify, extract
basicModel_neutral_lbs_10_207_0_v1.0.0.pkland rename it toSMPL_NEUTRAL.pkl. - Put the all the downloaded files into
./data/smplxfolder with following structure:
. └── data └── smplx ├── smpl │ └── SMPL_NEUTRAL.pkl └── smpl_uv.obj -
Download NeuMan dataset and pretrained models:
Alternately, run the following script to set up data and pretrained models.
bash setup_data_and_models.sh
-
(Optional) Download AMASS dataset for reposing:
- AMASS dataset is used for rendering novel poses, specifically
render_reposing.pyandrender_gathering.py. - We used SFU mocap(SMPL+H G) subset, please download from AMASS.
- Put the downloaded mocap data in to
./data/SFUfolder.
. └── data └── SFU ├── 0005 ├── 0007 │ ... └── 0018 - AMASS dataset is used for rendering novel poses, specifically
-
Render using pretrained model
Render 360 views of a canonical human:
python render_360.py --scene_dir ./data/bike --weights_path ./out/bike_human/checkpoint.pth.tar --mode canonical_360
Render 360 views of a posed human:
python render_360.py --scene_dir ./data/bike --weights_path ./out/bike_human/checkpoint.pth.tar --mode posed_360
Render test views of a sequence, and evaluate the metrics:
python render_test_views.py --scene_dir ./data/bike --weights_path ./out/bike_human/checkpoint.pth.tar
Render novel poses with the background:
python render_reposing.py --scene_dir ./data/bike --weights_path ./out/bike_human/checkpoint.pth.tar --motion_name=jumpandroll
Render telegathering:
python render_gathering.py --actors parkinglot seattle citron --scene_dir ./data/seattle --weights_path ./out/seattle_human/checkpoint.pth.tar
-
Download NeuMan dataset
-
Train scene NeRF
python train.py --scene_dir ./data/bike/ --name=bike_background --train_mode=bkg- Train human NeRF
python train.py --scene_dir ./data/bike --name=bike_human --load_background=bike_background --train_mode=smpl_and_offset- Preprocess: Check preprocess
@inproceedings{jiang2022neuman,
title={NeuMan: Neural Human Radiance Field from a Single Video},
author={Jiang, Wei and Yi, Kwang Moo and Samei, Golnoosh and Tuzel, Oncel and Ranjan, Anurag},
booktitle={Proceedings of the European conference on computer vision (ECCV)},
year={2022}
}
The code is released under the LICENSE terms.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent