inerf-pytorch implementation

This repository has been built over this pytorch implementation of NeRF: https://github.com/yenchenlin/nerf-pytorch

Dependencies:

- pytorch
- torchvision
- numpy
- imageio
- imageio-ffmpeg
- matplotlib
- configargparse
- opencv-python
- pyquaternion

Installation:

Run the following commands.

• cd inerf
• pip install -r requirements.txt

Implementation Details:

Currently, this implementation supports 2 datasets i.e. nerf-synthetic and nerf-llff. Performing pose estimation for any query image of a scene also requires its pretrained NeRF representation. This repository may not contain all the details to train NeRF. We have evaluated results for scenes lego (synthetic) and fern (llff).

run_inerf.py script performs pose estimation for a random query image from the validation/test dataset. The pose is randomly initialized in some vicinity of the ground truth camera pose. Please check out the report for more details.

It usually takes less than 5 minutes to optimize pose for 1 query image on a single GPU.

Data:

• lego (synthetic): pretrained NeRF, images and corresponding camera poses
• fern (llff): pretrained NeRF, images and corresponding camera poses

How to optimize camera pose: python run_inerf.py --config inerf_configs/{SCENE}.txt (replace {SCENE} with lego | fern)

Also, for more data, pretrained models and other implementation details on NeRF, checkout the file nerf_README.md

Code Citations:

• Pytorch NeRF implementation: https://github.com/yenchenlin/nerf-pytorch
• Modern Robotics for exponential parameterization: https://github.com/NxRLab/ModernRobotics/blob/master/packages/Python/modern_robotics/core.py