SG-NeRF

This is the official implement of ECCV 2024 paper "SG-NeRF: Neural Surface Reconstruction with Scene Graph Optimization".

Data Convention

The data is organized as follows:

<case_name>
|-- cameras_sphere.npz	    	# camera parameters
|-- database.db			# colmap database
|-- train			# input images from each view 
	|-- 000.png
	|-- 001.png
	...

Here the cameras_sphere.npz follows the data format in IDR, where world_mat_xx denotes the world to image projection matrix, and scale_mat_xx denotes the normalization matrix.

The database.db follows the database format in COLMAP.

Installation

pip install -r requirements.txt

Running

• Training on SG-NeRF dataset

  python exp_runner.py --case <case_name> --conf ./confs/sgnerf.conf --gpu <cuda_index>

• Training on DTU dataset

  python exp_runner.py --case <case_name> --conf ./confs/dtu.conf --gpu <cuda_index>

• Training on custom configuration

  python exp_runner.py --case <case_name> --conf <config_file> --gpu <cuda_index>

Acknowledgement

Some code snippets are borrowed from NeuS and COLMAP. Thanks for these great projects.