collector-m/mv3d_voxelnet

A combination of MV3D (from bostondiditeam) and VoxelNet (from qianguih)

Towards Better Point Cloud Feature Encoding in Multi-Input Neural Networks for Autonomous Driving Object Detection

Structure

This repository has two main folders, mv3d and voxelnet.

The first, mv3d, is based on MV3D (Multi-View 3D Object Detection Network), an autonomous driving model built with a region proposal network and a fusion network architecture that utilizes successive convolutional layers and residual connections. EDSR was first presented in a paper by Chen et al. that can be found here. The code in this folder is based off of a TensorFlow implementation of the paper that can be found here. The model's primary code and structure can be found in mv3d/src/mv3d_net.py.

The second, voxelnet, is based on VoxelNet, a model that focuses on constructing a feature representation of the LIDAR data. VoxelNet was first presented in a paper by Zhou and Tuzel that can be found here. The code in this folder is based off of a TensorFlow implementation of VoxelNet that can be found here. The model's primary code and structure can be found in the voxelnet/model folder.

Entering either folder will display the original GitHub's README file for each project.