Method

Fast and Accurate 3D Object Detection for Lidar-Camera-Based Autonomous Vehicles [la] [PFF3D]
https://ieeexplore.ieee.org/abstract/document/9340187

Submitted on 4 Sep. 2020 03:26 by
Lihua Wen (University of Ulsan)

Running time:0.05 s
Environment:GPU @ 3.0 Ghz (Python + C/C++)

Method Description:
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.
Parameters:
Loss function is the widely used in SECOND.
Latex Bibtex:
@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}}

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).


Benchmark Easy Moderate Hard
Car (Detection) 95.37 % 92.15 % 87.54 %
Car (Orientation) 94.86 % 91.06 % 86.28 %
Car (3D Detection) 81.11 % 72.93 % 67.24 %
Car (Bird's Eye View) 89.61 % 85.08 % 80.42 %
Pedestrian (Detection) 62.12 % 52.53 % 50.27 %
Pedestrian (Orientation) 48.75 % 40.99 % 38.99 %
Pedestrian (3D Detection) 43.93 % 36.07 % 32.86 %
Pedestrian (Bird's Eye View) 48.74 % 40.94 % 38.54 %
Cyclist (Detection) 79.44 % 66.25 % 60.11 %
Cyclist (Orientation) 78.02 % 64.06 % 58.06 %
Cyclist (3D Detection) 63.27 % 46.78 % 41.37 %
Cyclist (Bird's Eye View) 72.67 % 55.71 % 49.58 %
This table as LaTeX


2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot



2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot



2D object detection results.
This figure as: png eps pdf txt gnuplot



Orientation estimation results.
This figure as: png eps pdf txt gnuplot



3D object detection results.
This figure as: png eps pdf txt gnuplot



Bird's eye view results.
This figure as: png eps pdf txt gnuplot




eXTReMe Tracker