Learning Implicit Templates for Point-Based Clothed Human Modeling (ECCV 2022)

This repository contains the implementation of the paper

Learning Implicit Templates for Point-Based Clothed Human Modeling (ECCV 2022)

Siyou Lin, Hongwen Zhang, Zerong Zheng, Ruizhi Shao, Yebin Liu

Project page | Paper | Supp

A First-Implicit-Then-Explicit (FITE) pipeline for modeling humans in clothing.

[Update]

Uploaded the script for error measuring: measure_error.py. There was some typo in Eq. (6) and Eq. (7) in the the supplementary material. We used pymeshlab for measuring errors. The point-to-point distance is normalized by the bounding box diagonal when computing the Chamfer distance. Furthermore, both Chamfer distance and cosine similarity are divided by 2.

Overview

Our code consists of three main steps:

Given a minimal T-pose SMPL template, compute the diffused skinning field, stored as a 256^3 voxel grid.
Learn a canonical implicit template using SNARF, but with our diffused skinning field.
Render position maps (posmaps for short) and train the point-based avatars.

You can get a quick start with our pretrained models, or follow the instructions below to go through the whole pipeline.

Quick Start

To use our pretrained models (download from here), only the test script of step 3 needs to be executed.

First, unzip the downloaded pretrained data, put {subject_name}_clothed_template.npz in the folder data_templates, and put checkpoint-400.pt and geom-feats-400.pt in results/{resynth,cape}_pretrained/step3-checkpoints. Rename the check points to *-latest.pt

To animate a certain subject, say rp_carla_posed_004, prepare a pose sequence (.npz files containing the pose and transl parameters) under the directory data_scans/rp_carla_posed_004/test/, and then run (assuming the project directory is the working directory):

python -m step3_point_avatar.render_posmaps rp_carla_posed_004 test

This will render the posmaps to data_posmaps/rp_carla_posed_004/test. Then, run

python -m step3_point_avatar.test_fite_point_avatar

The animated point clouds will be saved at results/resynth_pretrained/step3-test-pcds.

The Whole Pipeline

Prior to running the pipeline, set expname in configs/common.yaml to the name of your experiment and keep it fixed.

1. Dependencies

We have tested this code on Ubuntu 20.04 with Python 3.8.10 and CUDA 11.1.

To run the whole pipeline, the user needs to install the following dependencies.

The PointInterpolant executable from https://github.com/mkazhdan/PoissonRecon. After successfully building PointInterpolant, set point_interpolant_exe in the config file configs/step1.yaml as the path to the executable.
```
git clone https://github.com/mkazhdan/PoissonRecon
cd PoissonRecon
make pointinterpolant
cd ..
```

Git clone https://github.com/NVIDIAGameWorks/kaolin.git to the FITE project directory and build it.

git clone --recursive https://github.com/NVIDIAGameWorks/kaolin
cd kaolin
git checkout v0.11.0     # optional, other versions should also work
python setup.py install 	# use --user if needed
cd ..

Download chamfer distance from this implementation: https://github.com/krrish94/chamferdist/tree/97051583f6fe72d5d4a855696dbfda0ea9b73a6a and build it.
```
cd chamferdist
python setup.py install		# use --user if needed
cd ..
```
Download the SMPL models from https://smpl.is.tue.mpg.de/ (v1.1.0). Put them in some folder structured as:
```
smpl_models/
  smpl/
    SMPL_MALE.pkl
    SMPL_FEMALE.pkl
    SMPL_NEUTRAL.pkl
```
Then set smpl_model_path in configs/common.yaml as the path to this folder.

Install these python packages:

pyyaml   # 6.0
tqdm     # 4.62.3
smplx    # 0.1.28
torch    # 1.10.0+cu111
trimesh  # 3.10.2
numpy    # 1.22.3
opencv-python  # 4.5.5
scikit-image   # 0.18.1
pytorch3d      # 0.6.1
pyglm    # 2.5.7
pyopengl # 3.1.0
glfw     # 2.1.0
scipy    # 1.8.0

2. Data Preparation for Training

To train on your own data, you need to prepare the scans (which must be closed meshes) as .npz files containing the following items:

'pose':		of shape (72,), SMPL pose parameters
'transl':	of shape (3,), translation of the scan
'scan_f':	of shape (N_faces, 3), triangle faces of the scan mesh
'scan_v':	of shape (N_vertices, 3), vertices of the scan mesh
'scan_pc':	of shape (N_points, 3), points uniformly sampled on the scan
'scan_n':	of shape (N_points, 3), normals of the sampled points (unit length)

Note that the number of points of scan_pc and scan_n must be the same across different scans (for tensor batching), while the number of vertices and faces can be different across scans. These .npz files should be placed at data_scans/{subject_name}/train.

A minimal SMPL body matching the scans is also needed. Prepare the T-pose SMPL mesh as {subject_name}_minimal_tpose.ply, and put it in data_templates/{subject_name}/ . Finally, add the gender of the subject to data_templates/gender_list.yaml.

3. Diffused Skinning

This step computes a voxel grid for diffused skinning. First, run

cd step1_diffused_skinning
sh compile_lbs_surf_grad.sh
cd ..

to compile the c++ program for computing surface gradient of LBS weights. Then, change the subject option in configs/step1.yaml to the name of your subject, and run

python -m step1_diffused_skinning.compute_diffused_skinning

After it finishes, a file named {subject_name}_cano_lbs_weights_grid_float32.npy will be placed at data_templates/{subject_name}/. You can optionally delete the intermediate files in data_tmp_constraints and data_tmp_skinning_grid if they take up too much space.

4. Implicit Templates

cd step2_implicit_templates
python setup.py install    # --user if needed
cd ..

Change datamodule.subject in configs/step2.yaml to the name of the subject to train. Then run

python -m step2_implicit_template.train_fite_implicit_template

Intermediate visualizations and checkpoints can be found at results/{expname}/step2-results and results/{expname}/step2-results. After the training is done, run

python -m step2_implicit_template.extract_fite_implicit_template

This extracts the implicit templates in canonical poses to data_templates.

5. Point Avatar

Prior to train the point avatar(s), the user needs to render the posmaps:

python -m step3_point_avatar.render_posmaps {subject_name} {dataset_split}

Recall that our point avatar part is a multi-subject model. If you do use multiple subjects, each should go through all the steps above. After that, collect the names and genders of all subjects in a config file configs/{expname}_subject_list.yaml. Then, run

python -m step3_point_avatar.train_fite_point_avatar

After training is complete, prepare test poses (only pose and transl are needed) in data_scans/{subject_name}/test/ and render posmaps same as above (test split). Then, run

python -m step3_point_avatar.test_fite_point_avatar

The outputted point clouds can be found at results/{expname}/step3-test-pcds/.

Acknowledgements & A Note on the License

This code is partly based on SNARF and POP. We thank those authors for making their code publicly available. Note that the code in step2_implicit_template is inherited from SNARF, while that in step3_point_avatar is herited from POP. Please follow their original licenses if you intend to use them.

jsnln / fite