GPU Sparse Voxel Octree

This repository contains a GPU-accelerated implementation of a sparse voxel octree (SVO) based on the techniques described in the papers "Fast Parallel Surface and Solid Voxelization on GPUs" (ACM) and "Data-Parallel Octrees for Surface Reconstruction" (IEEE). The project leverages CUDA for parallel processing to efficiently voxelized given 3D models and structure them into an SVO using Morton codes.

Key Features

• Surface Voxelization: Efficient GPU-based surface voxelization using the methodology outlined in cuda_voxelizer.
• Sparse Voxel Octree Construction: Constructs SVOs from voxelized data using a bottom-up approach with Morton codes for indexing and storage.
• Geometry Primitives Storage: Utilizes Morton codes to store non-redundant geometric primitives - vertices, edges, and faces of voxels.

Prerequisites

• NVIDIA GPU with CUDA Compute Capability 6.0 or higher.
• CUDA Toolkit 11.6 or higher.
• C++ compiler compatible with the CUDA version used. (C++17 at least)
• CMake(>=3.18) for building the project.

Setup

git clone https://github.com/Alan-Leo-Wong/GPU-SVO.git
cd GPU-SVO
mkdir build && cd build
cmake ..
cmake --build . -j your-core-number

Usage

After compiling the project, you can run the executable using:

./main <input-mesh.obj> <resolution>

• <input-mesh.obj>: Path to the input .obj file.
• <resolution>: Desired voxel resolution (e.g., 128).

The output will be an .obj file named <mesh_name>_depth_svo.obj, which represents the constructed sparse voxel octree at the specified resolution.