The KITTI Vision Benchmark Suite

This paper first presents an early-fusion method
to exploit both LiDAR and camera data for fast 3D
object detection with only one backbone,
achieving a good balance between accuracy and
efficiency. We propose a novel point feature
fusion module to directly extract point-wise
features from raw RGB images and fuse them with
their corresponding point cloud with no backbone.
In this paradigm, the backbone that extracts RGB
image features is abandoned to reduce the large
computation cost. Our method first voxelizes a
point cloud into a 3D voxel grid and utilizes two
strategies to reduce information loss during
voxelization.

Loss function is the widely used in SECOND.

@ARTICLE{9340187, author={L. -H. {Wen} and K. -H.
{Jo}}, journal={IEEE Access}, title={Fast and
Accurate 3D Object Detection for Lidar-Camera-Based
Autonomous Vehicles Using One Shared Voxel-Based
Backbone}, year={2021}, volume={9}, number={},
pages={22080-22089}, doi=
{10.1109/ACCESS.2021.3055491}}