Method

Patches (VoxelNets un-chained) [la] [Patches]
[Anonymous Submission]

Submitted on 28 Jan. 2019 17:53 by
[Anonymous Submission]

Running time:0.15 s
Environment:GPU @ 2.0 Ghz

Method Description:
1. A simpler VoxelNet without 3D-convolutions is
used as an RPN.
2. Patches, a mobile VoxelNet trained to operate
on small areas, reevaluates locations found by the
RPN.

trained with 50/50 split

v1.0 18.12.2018
3D - 87.07 - 76.56 - 68.65
BEV - 89.97 - 86.06 - 79.42
Parameters:
Voxel size RPN: 0.2 x 0.2 x 2.0 meters
Voxel size Patches: 0.15 x 0.15 x 4/19 meters
BEV size Patches: 9.6m x 9.6m

update v2.0: higher BEV resolution 0.2 -> 0.15
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) 90.75 % 89.61 % 87.42 %
Car (Orientation) 90.73 % 89.48 % 87.18 %
Car (3D Detection) 87.87 % 77.16 % 68.91 %
Car (Bird's Eye View) 89.78 % 86.55 % 79.22 %
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