This repository contains the Code release for my thesis, in which we use Neural Radiance Fields to perform Novel View Synthesis on unbounded dynamic urban scenes. We demonstrate our method on the Waymo Open Dataset and data we generate ourselves with CARLA. We combine approaches from Urban Radiance Fields (URF), Mip-NeRF360, Neural Scene Graphs (NSG) and BARF.
Our method is built on top of Mip-NeRF in JAX and in particular contains our own re-implementation of URFs LIDAR losses, Mip-NeRF360s spatial re-parameterization and BARFs frequency encoding filtering.
Modern self-driving systems collect vast amounts of data to train their advanced detection and decision-making algorithms. In contrast, simulation environments in which changes to a self-driving system are tested still rely on handcrafted assets and environment models in many cases. We propose a method that can use the data that is collected during normal operation of a self-driving car to model a scene in a Neural Radiance Field (NeRF), which allows us to perform Novel View Synthesis and view the scene from previously unseen camera perspectives.
To do this we have to overcome a few problems with modeling an unbounded outdoor scene in a NeRF, namely sampling efficiency in unbounded spaces, sparse views and dynamic scene content. We do this by combining advances from multiple recent state-of-the-art works. First we employ a non-linear scene re-parameterization to deal with sampling efficiency, which shrinks space so it is bounded again and can be easily sampled. Secondly we deal with sparse views by supervising the densities that are emitted by a NeRF with depth and volume carving losses. And lastly we decompose the scene into static background and dynamic parts with the help of 3D bounding box annotations and then train a separate NeRF for the background and each dynamic object, which allows us to manipulate camera and object pose separately.
Since we want to make this method easily scalable we relax the requirement for perfect human-annotated 3D bounding boxes and propose a method to optimise their position and orientation jointly with the radiance field. We accomplish this by framing the problem as a camera registration task and treating the transformation given by a 3D bounding box as an extrinsic camera matrix that needs to be registered with image data. Our method disables positional encoding to get around noisy gradients that can arise from backpropagating through it until the bounding boxes have converged and then uses anti-aliased integrated positional encoding to learn high frequency features.
We apply our method to a synthetic dataset which we generate with CARLA, a self-driving simulator, and the Waymo Open Dataset. Our method is capable of recovering reasonable 3D bounding boxes from errors up to half a meter from just five observations when we use LIDAR depth information in addition to RGB supervision. We also outperform a non-dynamic aware baseline in the Novel View Synthesis task for both ground truth and optimised bounding boxes.
Novel View Synthesis results from our method with rendered depth:
The results above are generated from a NeRF that was only trained on 25 Images and their corresponding depth images.
In another experiment we jointly optimise the bounding box poses and the radiance field:
The first gif shows our method when optimising the bounding box pose and the second without optimisation.
Clone this repo.
git clone https://github.com/FelTris/durf.git
cd durfCreate a new conda environment and depending on what you want to do install dependencies.
conda create -n durf python=3.6.13; conda activate durf
conda install pip; pip install --upgrade pip
# For waymo data pre-processing use requirements_wod, for carla use requirements_carla
pip install -r requirements_jax.txtInstall GPU support for Jax. If you run into trouble here, check for the official Jax installation guide.
# Remember to change cuda101 to your CUDA version, e.g. cuda110 for CUDA 11.0.
pip install --upgrade jax jaxlib==0.1.65+cuda101 -f https://storage.googleapis.com/jax-releases/jax_cuda_releases.htmlFor the waymo open data, download the sequences from the official website and run both the waymo_data and waymo_labels notebooks. To generate the sky masks use an off-the-shelf segmentation network and save the sky pixels as a mask. We also provide a few examples here for data we generate with CARLA. If you want an example of how the waymo data format is supposed to look, please contact me directly. I'm not sure how ok it is otherwise to share waymo open data publicly.
We provide example scripts for training with CARLA and Waymo data in scripts/.
Change the paths to where you put the data on your system.
Gin configuration files are provided for each dataset in configs/.
To evaluate or render your own trajectory we provide jupyter notebooks in notebooks/.
If you use this code in your research please cite:
@mastersthesis{durf22,
author = {Felix Tristram and
Matthias Niessner},
title = {Neural Rendering for Dynamic Urban Scenes},
school = {Technical University of Munich},
year = {2022},
note = {Visual Computing Group}
}