Method

3D Harmonic Loss: Towards Task-consistent and Time-friendly 3D Object Detection for V2X Orch [Harmonic PointPillar]
https://github.com/XJTU-Haolin/TT3D

Submitted on 14 Aug. 2022 16:48 by
Mekala M S (Yeungnam University)

Running time:0.01 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
We proposed 3D harmonic loss, inspired by related
works on solving inconsistency problem in image
domain, to alleviate the inconsistent predictions
in
LiDAR pointcloud. PointPillar was used as baseline
and our proposed method was applied.
Parameters:
Same parameters with PointPillar in mmdetection3D
codebase. Only difference is that Harmonic
PointPillar was trained via our proposed 3D
harmonic
loss.
Latex Bibtex:
@article{context,
title={3D Harmonic Loss: Towards Task-consistent
and Time-friendly 3D Object Detection for V2X
Orchestration},
author={Zhang, Haolin and Mekala, M S, Yang, Dongfang, John
Isaacs and Nain, Zulkar and Park, Ju H.
and Jung, Ho-Youl},
journal={will submit to IEEE Transactions on
Vehicular Technology},
year={2022},
publisher={IEEE}
}

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) 95.16 % 92.25 % 89.11 %
Car (Orientation) 94.23 % 90.78 % 87.42 %
Car (3D Detection) 82.26 % 73.96 % 69.21 %
Car (Bird's Eye View) 90.89 % 87.28 % 82.54 %
This table as LaTeX


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



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



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



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




eXTReMe Tracker