Method

VoxelNet(Unofficial) [VoxelNet(Unofficial)]


Submitted on 24 May. 2019 10:16 by
lin yan (Xidian University)

Running time:0.5 s
Environment:GPU @ 2.0 Ghz (Python)

Method Description:
This is an unofficial implementation of VoxelNet
from the
Paper `VoxelNet: End-to-End Learning for Point
Cloud Based
3D Object Detection` with TensorFlow and a single
GTX 1080Ti
card with 12GB memory.
Parameters:
batch_size=2
learning_rate=0.0005
training/validation split=1:1
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) 79.27 % 73.39 % 65.61 %
Car (Orientation) 33.54 % 30.29 % 27.36 %
Car (3D Detection) 74.59 % 64.80 % 57.38 %
Car (Bird's Eye View) 86.60 % 77.94 % 69.66 %
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