Method

Enhanced Frustum Pseudo LiDAR Framework for Monocular 3D Object Detection [ITS-MDPL]


Submitted on 1 Sep. 2020 06:31 by
ITS Lab Huawei Canada (Huawei Technologies Canada)

Running time:0.16 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
A frustum pseudo LiDAR framework for monocular 3D
object detection has been developed, where the
monocular depth estimation, instance segmentation
and 3D detection are enhanced with respect to the
objective of improving object detection accuracy.
Parameters:
TBD
Latex Bibtex:

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) 92.31 % 79.30 % 71.94 %
Car (Orientation) 92.05 % 78.24 % 70.73 %
Car (3D Detection) 23.81 % 14.21 % 12.11 %
Car (Bird's Eye View) 32.80 % 19.52 % 16.96 %
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




eXTReMe Tracker