How can I save mesh and texture files?

ARAH: Animatable Volume Rendering of Articulated Human SDFs

Paper | Project Page

This repository contains the implementation of our paper ARAH: Animatable Volume Rendering of Articulated Human SDFs.

You can find detailed usage instructions for using pretrained models and training your own models below.

If you find our code useful, please cite:

@inproceedings{ARAH:2022:ECCV,
  title = {ARAH: Animatable Volume Rendering of Articulated Human SDFs},
  author = {Shaofei Wang and Katja Schwarz and Andreas Geiger and Siyu Tang},
  booktitle = {European Conference on Computer Vision},
  year = {2022}
}

Installation

Environment Setup

This repository has been tested on the following platform:

Python 3.9.7, PyTorch 1.10 with CUDA 11.3 and cuDNN 8.2.0, Ubuntu 20.04/CentOS 7.9.2009

To clone the repo, run either:

git clone --recursive https://github.com/taconite/arah-release.git

or

git clone https://github.com/taconite/arah-release.git
git submodule update --init --recursive

Next, you have to make sure that you have all dependencies in place. The simplest way to do so, is to use anaconda.

You can create an anaconda environment called arah using

conda env create -f environment.yml
conda activate arah

Lastly, compile the extension modules. You can do this via

python setup.py build_ext --inplace

SMPL Setup

Download SMPL v1.0 for Python 2.7 from SMPL website (for male and female models), and SMPLIFY_CODE_V2.ZIP from SMPLify website (for the neutral model). After downloading, inside SMPL_python_v.1.0.0.zip, male and female models are smpl/models/basicmodel_m_lbs_10_207_0_v1.0.0.pkl and smpl/models/basicModel_f_lbs_10_207_0_v1.0.0.pkl, respectively. Inside mpips_smplify_public_v2.zip, the neutral model is smplify_public/code/models/basicModel_neutral_lbs_10_207_0_v1.0.0.pkl. Remove the chumpy objects in these .pkl models using this code under a Python 2 environment (you can create such an environment with conda). Finally, rename the newly generated .pkl files and copy them to subdirectories under ./body_models/smpl/. Eventually, the ./body_models folder should have the following structure:

body_models
 └-- smpl
    ├-- male
    |   └-- model.pkl
    ├-- female
    |   └-- model.pkl
    └-- neutral
        └-- model.pkl

Then, run the following script to extract necessary SMPL parameters used in our code:

python extract_smpl_parameters.py

The extracted SMPL parameters will be saved into ./body_models/misc/.

Quick Demo on the AIST++ Dataset

Run bash download_demo_data.sh to download and extract 1) pretrained models and 2) the preprocessed AIST++ sequence.
Run the pre-trained model on AIST++ poses via
```
python test.py --num-workers 4 configs/arah-zju/ZJUMOCAP-377-mono_4gpus.yaml
```
The script will compose a result .mp4 video in out/arah-zju/ZJUMOCAP-377-mono_4gpus/vis. There are a total of 258 frames, so it will take some time to render all the frames. If you want to check the result quickly run:
```
python test.py --num-workers 4 --end-frame 10 configs/arah-zju/ZJUMOCAP-377-mono_4gpus.yaml
```
to render only the first 10 frames, or
```
python test.py --num-workers 4 --subsampling-rate 25 configs/arah-zju/ZJUMOCAP-377-mono_4gpus.yaml
```
to render every 25th frame. Inference requires ~20GB VRAM, if you don't have so much memory, add --low-vram option. This should run with ~12GB VRAM at the cost of longer inference time.

Results on ZJU-MoCap

For easy comparison to our approach, we also store all our rendering and geometry reconstruction results on the ZJU-MoCap dataset here. Train/val splits on cameras/poses follow NeuralBody's split. Pseudo ground truths for geometry reconstruction on the ZJU-MoCap dataset are stored in this folder. For evaluation script and data split of geometry reconstruction please refer to this comment.

Dataset preparation

Due to license issues, we cannot publicly distribute our preprocessed ZJU-MoCap and H36M data. You have to get the raw data from their respective sources and use our preprocessing script to generate data that is suitable for our training/validation scripts. Please follow the steps in DATASET.md.

Download pre-trained skinning and SDF networks

We provide pre-trained models on the CAPE dataset as prerequisites, including 1) meta learned skinning network on the CAPE dataset, 2) MetaAvatar SDF model. After downloading them, please put them in respective folders under ./out/meta-avatar.

Training

To train new networks from scratch, run

python train.py --num-workers 4 ${path_to_config}

Where ${path_to_config} is the relative path to the yaml config file, e.g. config/arah-zju/ZJUMOCAP-313_4gpus.yaml

Training and validation use wandb for logging, which is free to use but requires online register.

Note that by default, all models are trained on 4 GPUs with a total batch size of 4. You can change the value of training.gpus to [0] in the configuration file to train on a single GPU with a batch size of 1, however the model accuracy may drop and the training might become less stable.

Pre-trained models of ARAH (Work In Progress)

We provide pre-trained models, including multi-view and monocular models. After downloading them, please put them in respective folders under ./out/arah-zju or ./out/arah-h36m.

Validate the trained model on within-distribution poses

To validate the trained model on novel views of training poses, run

python validate.py --novel-view --num-workers 4 ${path_to_config}

To validate the trained model on novel views of unseen poses, run

python validate.py --novel-pose --num-workers 4 ${path_to_config}

Test the trained model on out-of-distribution poses

To run the trained model on preprocessed poses, run

python test.py --num-workers 4 --pose-dir ${pose_dir} --test-views ${view} configs/arah/${config}

where ${pose_dir} denotes the directory under data/odp/CoreView_${sequence_name}/ that contains target (out-of-distribution) poses. ${view} indicates the testing views from which to render the model.

Currently, the code only supports animating ZJU-MoCap models for out-of-distribution poses.

License

We employ MIT License for the ARAH code, which covers

extract_smpl_parameters.py
train.py
validate.py
test.py
setup.py
configs
im2mesh/
preprocess_datasets/preprocess_ZJU-MoCap.py

Our SDF network is based on SIREN. Our mesh extraction code is borrowed from DeepSDF. The structure of our rendering code is largely based on IDR. Our root-finding code is modified from SNARF. We thank authors of these papers for their wonderful works which inspired this paper.

Modules not covered by our license are:

Modified code from EasyMocap to preprocess ZJU-MoCap/H36M datasets (./preprocess_datasets/easymocap);
Modified code from SMPL-X (./human_body_prior);

for these parts, please consult their respective licenses and cite the respective papers.

	curr_start_points_opt, acc_start_dis_opt, curr_start_transforms_opt, converge_mask_opt = \
	search_iso_surface_depth(cam_locs,
	body_ray_directions,
	~diverge_mask if eval else torch.ones_like(diverge_mask),
	curr_start_points,
	acc_start_dis,
	curr_start_transforms,
	sdf_network,
	loc,
	sc_factor,
	skinning_model,
	vol_feat,

	# Get bone transforms
	body = body_model(root_orient=new_root_orient_torch, pose_body=pose_body_torch, pose_hand=pose_hand_torch, betas=betas_torch, trans=new_trans_torch)

	body_model_em = load_model(gender='neutral', model_type='smpl')
	verts = body_model_em(poses=poses_torch, shapes=betas_torch, Rh=new_root_orient_torch, Th=new_trans_torch, return_verts=True)[0].detach().cpu().numpy()

	vertices = body.v.detach().cpu().numpy()[0]
	new_trans = new_trans + (verts - vertices).mean(0, keepdims=True)
	new_trans_torch = torch.from_numpy(new_trans).cuda()

taconite / arah-release Goto Github PK

arah-release's Introduction

ARAH: Animatable Volume Rendering of Articulated Human SDFs

Paper | Project Page

Installation

Environment Setup

SMPL Setup

Quick Demo on the AIST++ Dataset

Results on ZJU-MoCap

Dataset preparation

Download pre-trained skinning and SDF networks

Training

Pre-trained models of ARAH (Work In Progress)

Validate the trained model on within-distribution poses

Test the trained model on out-of-distribution poses

License

arah-release's People

Contributors

Stargazers

Watchers

Forkers

arah-release's Issues

Recommend Projects

Recommend Topics

Recommend Org