Skip to content

[ECCV 2024] MetaCap: Meta-learning Priors from Multi-View Imagery for Sparse-view Human Performance Capture and Rendering

Notifications You must be signed in to change notification settings

sunshinnnn/metacap

Folders and files

NameName
Last commit message
Last commit date

Latest commit

Β 

History

3 Commits
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 
Β 

Repository files navigation

MetaCap

This repository contains the implementation of our ECCV 2024 paper "MetaCap: Meta-learning Priors from Multi-View Imagery for Sparse-view Human Performance Capture and Rendering". You can visit project page for more illustrations.

πŸ› οΈ Requirements

If GCC<11.0:

  • Create conda environment conda create --name metacap python=3.8
  • Install cuda11.6: conda install cudatoolkit=11.6 cudnn -c conda-forge, then conda install cudatoolkit-dev=11.6 -c conda-forge
  • Install PyTorch==1.13.1+cu116: pip install torch==1.13.1+cu116 torchvision==0.14.1 --extra-index-url https://download.pytorch.org/whl/cu116
  • Install tiny-cuda-nn PyTorch extension: pip install git+https://github.com/NVlabs/tiny-cuda-nn/@v1.6#subdirectory=bindings/torch
  • pip install -r requirements.txt
  • Install pytorch3d: pip install --no-cache-dir pytorch3d==0.7.2 -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu116_pyt1131/download.html
  • Install kaolin: pip install kaolin==0.13.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-1.13.1_cu116.html
  • Install nerfacc: pip install nerfacc==0.3.3 -f https://nerfacc-bucket.s3.us-west-2.amazonaws.com/whl/torch-1.13.0_cu116.html
  • Install point inside-outside check module:
    cd ./tools/extensions/libmesh
    python setup.py build_ext --inplace
    cd ../../..
    
If GCC>=11.0:
  • Create conda environment conda create --name metacap python=3.9
  • Install cuda11.8: conda install cudatoolkit=11.8 cudnn -c conda-forge, cudatoolkit-dev=11.8 doesn't exist yet.
  • Install PyTorch==2.0.0+cu118: pip install torch==2.0.0 torchvision==0.15.1 torchaudio==2.0.1 --index-url https://download.pytorch.org/whl/cu118 --no-cache-dir
  • Install tiny-cuda-nn 1.7 PyTorch extension from source: https://github.com/NVlabs/tiny-cuda-nn?tab=readme-ov-file#compilation-windows--linux
  • pip install -r requirements.txt
  • Install pytorch3d: pip install --no-cache-dir pytorch3d==0.7.3 -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py39_cu118_pyt200/download.html
  • Install kaolin: pip install kaolin==0.14.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.0.0_cu118.html
  • Install nerfacc 0.3.3 from source: https://github.com/nerfstudio-project/nerfacc/releases
  • Install point inside-outside check module:
    cd ./tools/extensions/libmesh
    python setup.py build_ext --inplace
    cd ../../..
    

πŸŽ₯ Data

Prepare Data
# Step 1: Download demo data from https://gvv-assets.mpi-inf.mpg.de/MetaCap and place them in ./datas
# datas
#   β”œβ”€β”€ ddc_configs (exists there)
#   β”œβ”€β”€ vis (exists there)
#   β”œβ”€β”€ preprocess_data.py (exists there)
#   β”œβ”€β”€ render_depth_s5.py (exists there)
#   β”œβ”€β”€ color_proxy_s2
#   β”œβ”€β”€ rotate_dir
#   └── demo

# Step 2: Download SMPL from: https://smpl.is.tue.mpg.de/ and place the model in ./datas/SMPLX/smpl/
# datas
#   └── SMPLX/smpl/
#      Β Β  β”œβ”€β”€ SMPL_FEMALE.pkl
#      Β Β  β”œβ”€β”€ SMPL_MALE.pkl
#      Β Β  └── SMPL_NEUTRAL.pkl

# Step 3: Download WildDynaCap and DynaCap from: https://gvv-assets.mpi-inf.mpg.de/MetaCap to /XX/XX/WildDynaCap_official(where to save dataset)
#         Edit the script "./data/process_data.py", base_dir = /XX/XX/WildDynaCap_official
#         Run the following scripts to generate training datas and testing datas:
cd ./datas
python ./preprocess_data.py -subi 2 -ct tight -st training
python ./preprocess_data.py -subi 2 -ct tight -st training -cs
python ./preprocess_data.py -subi 2 -ct tight -st testing
python ./preprocess_data.py -subi 2 -ct tight -st testing -cs
python ./preprocess_data.py -subi 2 -ct tight -st inTheWild
python ./preprocess_data.py -subi 2 -ct tight -st inTheWild -cs
ln -s /XX/XX/WildDynaCap_official/Subject0002 ./Subject0002
# change the subj idx and ct to process other subjects
cd ..

🚩 Edit Config before Training or Testing

  1. Edit configs/base.yaml root_dir for where you put this repository.
  2. Edit exp_dir and runs_dir for where to save experiment results and logs.
  3. Link result and log folders to under the repository folder. ln -s /XXX/XXX/results/metacap/exp ./exp, and ln -s /XXX/XXX/results/metacap/runs ./runs
Config Flag Meaning
dataset.cano_motion: template or cano, the motion of canonical pose.
dataset.deformer: ddc or smpl, the human template used in experiments.
dataset.rotate: True or False, spherial rendering or specific view rendering during testing phase.
dataset.smoothDQ: True or False, smooth the DQ format transformations or not.
dataset.preload: True or False, preload images, masks or not during the meta-learning phase.

model.grid_pre: True or False, pre-compute grid density or not.
model.decay_step: The iteration step to decay threshold_outer during training phase.

export.save_mesh: True or False, saveing mesh or not during testing or evaluation phase.

πŸ† Fast Demo

Demo in the Dome
# Run finetuning on sparse dome images and visualize them through spherical rendering.
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta-demo.yaml --gpu 0 --train \
        --ckpt  ./datas/demo/Subject0002/tight/ckpt/epoch=0-step=2000.ckpt \
        dataset.frame=1000 tag=Debug_Cano_Meta_DDC_Rotate dataset.cano_motion=template \
        trainer.val_check_interval=4000 dataset.threshold_outer=0.05 model.decay_step=5000 dataset.rotate=True
Demo in the Wild
# Run finetuning on sparse itw images and visualize them through spherical rendering.
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-itw-finetune2-new2-meta-demo.yaml --gpu 0 --train \
        --ckpt  ./datas/demo/Subject0002/tight/ckpt/epoch=0-step=2000.ckpt \
        dataset.frame=3300 tag=Debug_Cano_Meta_DDC_Rotate dataset.cano_motion=template \
        trainer.val_check_interval=4000 dataset.threshold_outer=0.05 model.decay_step=5000 dataset.rotate=True

πŸƒ Run

Meta-learning Prior
# Meta-learning with DDC
python launch.py --config configs/s2/metalearning/metaneusseq-domedenserawseq_meta-ddc-s2-smooth2-24-more-newthreshold0-100Frame-100Camera.yaml --gpu 0 --train \
tag=Test_Down2.0_Blur_100Frame_100Camera_DDC dataset.img_downscale=2.0 dataset.blur=True dataset.preload=True
# Meta-learning with SMPL
python launch.py --config configs/s2/metalearning/metaneusseq-domedenserawseq_meta-ddc-s2-smooth2-24-more-newthreshold0-100Frame-100Camera.yaml --gpu 0 --train \
tag=Test_Down2.0_Blur_100Frame_100Camera_SMPL dataset.img_downscale=2.0 dataset.blur=True dataset.preload=True dataset.deformer=smpl dataset.threshold_outer=0.075
Fine-tuning on Sparse-view Images
# Run finetuning and visualize them through spherical rendering.
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta.yaml --gpu 0 --train \
        --ckpt  ./exp/Subject0002/metaneusseq-domedenseseq_meta-Subject0002/xxxxxxxxxxxxxxxx/ckpt/epoch=0-step=2000.ckpt \
        dataset.frame=3000 tag=Cano_Meta_DDC_Rotate dataset.cano_motion=template \
        trainer.val_check_interval=4000 dataset.threshold_outer=0.05 model.decay_step=5000 dataset.rotate=True
# Run finetuning and run test code for given camera parameters.
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta.yaml --gpu 0 --train \
        --ckpt  ./exp/Subject0002/metaneusseq-domedenseseq_meta-Subject0002/xxxxxxxxxxxx/ckpt/epoch=0-step=2000.ckpt \
        dataset.frame=3000 tag=Cano_Meta_DDC dataset.cano_motion=template \
        trainer.val_check_interval=4000 dataset.threshold_outer=0.05 model.decay_step=5000 dataset.rotate=False
# Fine-tuning with SMPL
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta.yaml --gpu 0 --train \
        --ckpt  ./exp/Subject0002/metaneusseq-domedenseseq_meta-Subject0002/xxxxxxxxxxxxxxxx-SMPL/ckpt/epoch=0-step=2000.ckpt \
        dataset.frame=3000 tag=Cano_Meta_SMPL dataset.cano_motion=template dataset.deformer=smpl \
        trainer.val_check_interval=4000 dataset.threshold_outer=0.075 model.decay_step=5000 dataset.rotate=True
Fine-tuning on Monocular Image with Occlusion Handling
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta-one-oh.yaml --gpu 0 --train \
       --ckpt  ./exp/Subject0002/metaneusseq-domedenseseq_meta-Subject0002/xxxxxxxxxxxxxxxx/ckpt/epoch=0-step=2000.ckpt \
       dataset.frame=5200 tag=Cano_Meta_DDC_One_OH dataset.cano_motion=template dataset.deformer=ddc trainer.val_check_interval=4000 \
       dataset.threshold_ddc=0.05 model.decay_step=5000 checkpoint.every_n_train_steps=5000 dataset.rotate=True model.grid_pre=False dataset.compute_occlusion=True
Training on Dense-view Images for Ground-truth Geometry (Random Initialization)
python launch.py --config configs/s2/groundtruth/neus-domedenseraw-ddc-s2-smooth2-test-gt.yaml  --gpu 0 --train \
        dataset.frame=3000 tag=World_GT dataset.cano_motion=world trainer.val_check_interval=4000 \
        dataset.threshold_outer=0.1 dataset.threshold_ddc=0.1  model.decay_step=5000 dataset.smoothDQ=False
Fine-tuning FPS Evaluation (Testing on NVIDIA RTX 3090, >20FPS)
# We found the first time run will be a bit slower, run multiple times to get the stable speed.
python launch.py --config configs/s2/finetuning/neus-domedenseraw-ddc-s2-smooth2-test-finetune2-new2-meta.yaml --gpu 0 --train --notest \
      --ckpt ././exp2/metaneusseq-domedenseseq-Subject0002/XXXXXXXXXXXXXXX/ckpt/epoch=0-step=2000.ckpt\
       dataset.frame=5200 tag=Debug_Time dataset.cano_motion=template dataset.deformer=ddc trainer.val_check_interval=4000 \
       dataset.threshold_ddc=0.05 model.decay_step=5000 checkpoint.every_n_train_steps=5000 dataset.rotate=False model.grid_pre=True export.save_mesh=False  
Interpolation on Two Fine-tuned Frames and Rendering in Canonical Space
python launch.py --config configs/s2/interpolation/neusinterpolation-domedenseraw-ddc-s2-smooth2-test.yaml --gpu 0 --test \
    --ckpt1 ./exp/Subject0002/XXXXXX/ckpt/epoch=0-step=3000.ckpt \
    --ckpt2 ./exp/Subject0002/XXXXXX/ckpt/epoch=0-step=3000.ckpt  \
    dataset.frame=XXXXX tag=Debug_Interpolation dataset.rotate=True \
    dataset.rotate_template=True dataset.rotate_scale=0.0 export.save_mesh=False 

πŸ“ Evaluate

Evaluation scripts
cd ./evaluations
# prepare GT results and MetaCap results
python ./00_move_Mesh_GT.py
python ./00_prepare_T_w2s.py
python ./00_move_RGB_Mask_GT.py
python ./00_move_metacap.py
# evaluate rendering and geometry results
python ./01_evaluate_rgb_metacap.py
python ./01_evaluate_mesh_metacap.py
# summarize the final results
python ./02_summary_evaluate_metacap.py

πŸ“‘ Code Structure

Details
# Configs
- config files for meta-learning, fine-tuning, ground truth generation and interpolation experiments.

# Datas
- render_depth_s5.py: render depth maps for s5 if needed it during the prior training.
- preprocess_data.py: extract cropped images, masks and camera parameters from raw videos.
- color_proxy_s2: pre-computed proxy images for the occlusion handling strategy.
- ddc_configs: configs files for the DDC template and skel_tools.py
- SMPLX: datas for the SMPL template.
- rotate_dir: some placeholders for the spherical rendering. To be removed in the future version.

# Datasets
- domedenseraw.py: fine-tuning or ground-truth fitting or weight interpolation inference on single frame.
- domedenserawseq_meta.py: meta-learning on multiple frames.
- human_info.py: information of meta-learning views and frames, testing views and frames. Note that the motion and images of training split and testing split are not overlapped.

# Evaluations
- 00_move_xxx.py: moving results to a new folder for evaluation.
- 01_evaluate_xxx.py evaluating mesh and image metrics.
- 02_summary_xxx.py summarize the evaluation results.

# Models
- xxx.py: basic models for radiance fileds learning and feature embedding.

# Systems
- xxx.py: basic systems for model training, testing, and evaluation.

# Tools
- skel_tools.py: skeleton, skinned character and embedded graph used in DDC template.
- pyrender_tools.py: tools for skeleton, mesh, ground and axis visualization.
- cam_tools.py: functions for camera parameters loading.
- metric_tools.py: functions for geometry and image evaluation.
- omni_tools.py: utils for resizing, file creating, file copying...
- mesh_tools.py: utils for mesh loading, mesh saving and point2mesh prjections...

🌹 Acknowledgements

Some great prior works and codes we benefit from: instant-nsr-pl, NerfAcc, PyTorch3D, Kaolin, GRAB, InstantAvatar, EasyMocap, DDC.

πŸ“– License

Both the source code and our data are under the terms of the Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license. This project is only for research or education purposes, and not freely available for commercial use or redistribution.

🍺 Citation

If you find the codes or datasets from this repository helpful in your projects, welcome to cite the papers and give a star.

@inproceedings{sun2024metacap,
title = {MetaCap: Meta-learning Priors from Multi-View Imagery for Sparse-view Human Performance Capture and Rendering},
author = {Sun, Guoxing and Dabral, Rishabh and Fua, Pascal and Theobalt, Christian and Habermann, Marc},
year = {2024},
booktitle={ECCV}
}
@article{habermann2021,
author = {Marc Habermann and Lingjie Liu and Weipeng Xu and Michael Zollhoefer and Gerard Pons-Moll and Christian Theobalt},
title = {Real-time Deep Dynamic Characters},
journal = {ACM Transactions on Graphics}, 
month = {aug},
volume = {40},
number = {4}, 
articleno = {94},
year = {2021}, 
publisher = {ACM}
} 

About

[ECCV 2024] MetaCap: Meta-learning Priors from Multi-View Imagery for Sparse-view Human Performance Capture and Rendering

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages