Method

Hongyu-Net[la] [HNet-3DSSD]
[Anonymous Submission]

Submitted on 24 May. 2021 20:28 by
[Anonymous Submission]

Running time:0.05 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
Abstract we present a lightweight point-based 3D
single stage object detector using point cloud
density information to achieve decent balance of
accuracy and efficiency. Our approach improve the
detector robustness by using the DBSCAN cluster
mehod to get the cluster of the unordered point
cloud, then filter out uesless points by cluster
information. We propose a new Set Abstract(SA)
module in downsampling process to make detection
on less representative points feasible and
harmonize the high-level presentation feature and
low-level space feature. The new model including
an anchor-free regression head with a 3D
centerness assignment strategy, is developed to
meet the demand of high accuracy and speed.
Density-Net demonstrates accuracy and run-time
performance on par with the well-established and
highly-optimized 3DSSD baseline.
Parameters:
\shrink=[0,1,2]
Latex Bibtex:
None

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) 94.91 % 91.35 % 87.47 %
Car (Orientation) 0.01 % 0.47 % 0.63 %
Car (3D Detection) 86.06 % 76.48 % 69.71 %
Car (Bird's Eye View) 91.65 % 86.69 % 81.05 %
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