NeRF

This is an implementation of NeRF. Paper

Getting Started

Installation

- Clone this repo:

git clone https://github.com/samcao416/NeRF_vrvc.git
cd NeRF_vrvc

- Install PyTorch and other dependencies using:

conda create -n nerf python=3.8.5
conda activate nerf    
conda install pytorch==1.9.0 torchvision torchaudio cudatoolkit=11.1 -c pytorch -c conda-forge
conda install imageio
pip install yacs kornia
pip install imageio-ffmpeg
pip install tensorboard==2.2.0

Datasets

The synthetic datasets can be downloaded from here.

To Run the Program

- In the terminal, mkdir configs, then cp ./config_example/hotdog.yml /configs. - Change the training directory(DATASETS:TRAIN) and output directory(OUTPUT_DIR) as you want. - cd tools and run python train_net.py -c ../configs/hotdog.yml

To Monitor the trainning process

- After you can see the PSNR and loss, you can start a new terminal, and cd [OUTPUT_DIR] . - Run tensorboard --logdir . - Now you can watch the PSNR and loss curve as well as the training images in the webpage.

To Render New Views

After the the first checkpoint is saved to the output folder, you can render new views.
cd tools and run python renderer.py -c ../configs/hotdog.yml -t linear
Now two types of rendering are provided: gt and linear, the latter is the default option. gt stands for rendering the ground truth view images. linear stands for generating new camera poses based on the ground truth views by linear interpolation. By default, two novel views will be interpolated between two consecutive poses. However, the process of sorting ground truth cameras is still under construction. Before that, you'd better sort the cameras by yourself, or the novel views might be disordered (while the results will still be correct).

Acknowlegements

We borrowed some codes from Multi-view Neural Human Rendering (NHR).

samcao0416/NeRF_VRVC