wangyida / neural-actor

Rendering, Animating and Meshing Actors with NeRF

Rendering, Animating and Meshing Actors with NeRF

A library for rendering neural actors, and benchmarking dynamic NeRF

Examplar

Point Cloud Completion

Neural Rendering

Dynamic Meshing

SMPL Fitting

Setup

Notice that working on NVidia 3090 requires certain pytorch and torchvision versions

pip install torch==1.11.0 torchvision==0.12.0 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113

Or download the wheel and install with in case pytorch site is not included in the pip source

pip install torch-1.11.0+cu113-cp37-cp37m-linux_x86_64.whl

As for COLMAP, version 3.8 (the lastest in early 2023) has problen with CUDA 11.X, version 3.7 fits the best for SIFT feature extraction and matching with CUDA.https://github.com/colmap/colmap/tree/3.7

Blender

/snap/bin/blender ~/Downloads/rp_aliyah_4d_004_dancing_BLD/rp_aliyah_4d_004_dancing_2k.blend --background --python 'render/renderpeople.py' -- --with_images --start 1 --end 150

Format the camera parameters

python3 format_convert/json2yaml.py ~/Documents/datasets

Then add %YAML:1.0 as header for the newly generated yml file so that OpenCV can parse it

:%s/-\ \ //g
:%s/\ \ \ \ \ \ /\ \ \ /g

Mocap

Detect hand and face keypoints using OpenPose

python3 scripts/preprocess/extract_video.py ~/Documents/datasets --openpose ~/Downloads/openpose --handface --ext png --with_img --end 1

python3 apps/demo/mv1p.py ~/Documents/datasets --out ~/Documents/datasets/output/smpl --vis_det --vis_repro --undis --vis_smpl --end 1

Citation

@misc{rama2023wang,
Author = {Yida Wang},
Year = {2023},
Note = {https://github.com/wangyida/neural-actor},
Title = {Rendering, Animating and Meshing Actors with NeRF}
}