Stereo Radiance Fields

Julian Chibane, Aayush Bansal, Verica Lazova, Gerard Pons-Moll

Stereo Radiance Fields (SRF): Learning View Synthesis for Sparse Views of Novel Scenes

In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2021


A linux system with python environment manager conda and a full and system wide installation of the CUDA Toolkit 10.1 is required for the project (the latter only for compilation of the torchsearchsorted library).

The following commands clone the repo on your machine and install an environment, "srf", containing all dependencies.

git clone ADD LINK
cd SRF_git
conda env create -f srf_env.yml

Please close the terminal session at this point, and reopen it at the same location. This is done to ensure conda correctly loads all packages.

conda activate srf
pip install torchsearchsorted/

Data Setup

With the next commands the DTU MVS dataset is downloaded and put in place.

wget -P data/
unzip data/ -d data/
mv data/Rectified/* data/DTU_MVS
rmdir data/Rectified

Quick Start with Pretrained Model

To synthesise novel views use the following command

python --config configs/finetune_scan23.txt --video --render_factor 8 --generate_specific_samples scan23 --fixed_batch 1 --ft_path checkpoint.tar --gen_pose 0

where --config specifies the path to the experiment configuration and --gen_pose is the frame number from 0-55 (including both).