[ICCV 2023] Urban Radiance Field Representation with Deformable Neural Mesh Primitives

Introduction

The repository contains the official implementation of source code and pre-trained models of our paper:"Urban Radiance Field Representation with Deformable Neural Mesh Primitives". It is a new representation to model urban scenes for efficient and high-quality rendering!

Updates

2023.07.21: The:star::star::star:source code⭐⭐⭐is released! Try it!
2023.07.21: The:fire::fire::fire:pre-print🔥🔥🔥is released! Refer to it for more details!
2023.07.19: The project page is created. Check it out for an overview of our work!

Datasets

We conduct experiments on two outdoor datasets: KITTI-360 dataset, Waymo-Open-Dataset. Please refer to preprocess/README.md for more details.

Environments

Compile fairnr.

python setup.py build_ext --inplace

Main requirements:

CUDA (tested on cuda-11.1)
PyTorch (tested on torch-1.9.1)
pytorch3d
torchsparse

Other requirements are provided in requirements.txt

Training

Optimize geometry using our pre-trained auto-encoder by running sh scripts/train_${DATASET}_geo.sh. (Please specify SEQUENCE,DATA_ROOT,LOG_DIR and CKPT_DIR in the script.)
Train radiance field by running sh scripts/train_${DATASET}_render.sh. (Please specify SEQUENCE,DATA_ROOT,LOG_DIR, CKPT_DIR and PRETRAINED_GEO in the script.)

Evaluation

You can run scripts/test_kitti360.sh for evaluation. (Please specify SAVE_DIR, DATA_ROOT and the pretrained files in the script.)

To-Do List

Release Code and pretrained model
Technical Report
Project page

Citation

If you find this project useful for your work, please consider citing:

@article{lu2023dnmp,
  author    = {Lu, Fan and Xu, Yan and Chen, Guang and Li, Hongsheng and Lin, Kwan-Yee and Jiang, Changjun},
  title     = {Urban Radiance Field Representation with Deformable Neural Mesh Primitives},
  journal   = {ICCV},
  year      = {2023},
}

TurtleZhong / DNMP