MixVoxels: Mixed Neural Voxels for Fast Multi-view Video Synthesis

Project Page | Paper

Pytorch implementation for the paper: Mixed Neural Voxels for Fast Multi-view Video Synthesis.

cat.mp4

We present MixVoxels to better represent the dynamic scenes with fast training speed and competitive rendering qualities. The proposed MixVoxels represents the 4D dynamic scenes as a mixture of static and dynamic voxels and processes them with different networks. In this way, the computation of the required modalities for static voxels can be processed by a lightweight model, which essentially reduces the amount of computation, especially for many daily dynamic scenes dominated by the static background. As a result, with 15 minutes of training for dynamic scenes with inputs of 300-frame videos, MixVoxels achieves better PSNR than previous methods.

Installation

Install environment:

conda create -n mixvoxels python=3.8
conda activate mixvoxels
pip install torch torchvision
pip install tqdm scikit-image opencv-python configargparse lpips imageio-ffmpeg kornia lpips tensorboard pyfvvdp

Dataset

Download the Plenoptic Video Dataset
Unzip to your directory DATADIR and run the following command:

python tools/prepare_video.py ${DATADIR}

Training

To train a dynamic scene, run the following commands, you can train different dynamic scenes by assign DATA to different scene name:

DATA=coffee_martini # [coffee_martini|cut_roasted_beef|cook_spinach|flame_salmon|flame_steak|sear_steak]
# MixVoxels-T
python train.py --config configs/schedule5000/${DATA}_5000.txt --render_path 0
# MixVoxels-S
python train.py --config configs/schedule7500/${DATA}_7500.txt --render_path 0
# MixVoxels-M
python train.py --config configs/schedule12500/${DATA}_12500.txt --render_path 0
# MixVoxels-L
python train.py --config configs/schedule25000/${DATA}_25000.txt --render_path 0

Please note that in your first running, the above command will first pre-process the dataset, including resizing the frames by a factor of 2 (to 1K resolution which is a standard), as well as calculating the std of each video and save them into your disk. The pre-processing will cost about 2 hours, but is only required at the first running. After the pre-processing, the command will automatically train your scenes.

We provide the trained model:

scene	PSNR		download
scene	MixVoxels-T (15min)	MixVoxels-M (40min)	MixVoxels-T (15min)	MixVoxels-M (40min)
coffee-martini	28.1339	29.0186	link	link
flame-salmon	28.7982	29.2620	link	link
cook-spinach	31.4499	31.6433	link	link
cut-roasted-beef	32.4078	32.2800	link	link
flame-steak	31.6508	31.3052	link	link
sear-steak	31.8203	31.2136	link	link

Rendering and Generating Spirals

The following command will generate 120 novel view videos, or you can set the render_path as 1 in the above training command.

python train.py --config your_config --render_only 1 --render_path 1 --ckpt log/your_config/your_config.ckpt

Generating spirals:

python tools/make_spiral.py --videos_path log/your_config/imgs_path_all/ --target log/your_config/spirals --target_video log/your_config/spirals.mp4

Citation

If you find our code or paper helps, please consider citing:

@article{wang2022mixed,
  title={Mixed Neural Voxels for Fast Multi-view Video Synthesis},
  author={Wang, Feng and Tan, Sinan and Li, Xinghang and Tian, Zeyue and Liu, Huaping},
  journal={arXiv preprint arXiv:2212.00190},
  year={2022}
}

Acknowledge

The codes are based on TensoRF, many thanks to the authors.

johnnylu305 / mixvoxels