Method

Image Guidance Region Proposal [IGRP]
[Anonymous Submission]

Submitted on 4 Apr. 2020 03:53 by
[Anonymous Submission]

Running time:0.18 s
Environment:1 core @ 2.5 Ghz (Python + C/C++)

Method Description:
We leverage 2D detection from RGB image to help
generate 3D proposals from LiDAR point cloud,
which 2D bounding boxes are used to choose the
candidates whose centers located in the boxes.
Then, instead of sphere space, we apply cylinder
space to search neighbor point for learning
representation of point cloud in PointNet++.
Finally, we attach an interest over union
prediction branch to weight classification for
proposals refinement.
Parameters:
\rho = 0.5
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) 96.28 % 92.78 % 87.81 %
Car (Orientation) 96.27 % 92.66 % 87.63 %
Car (3D Detection) 86.27 % 75.90 % 69.31 %
Car (Bird's Eye View) 92.04 % 86.21 % 81.30 %
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