Method

Semantics-Augmented Set Abstraction [la] [SASA]
https://github.com/blakechen97/SASA

Submitted on 22 Feb. 2022 04:43 by
Chen Chen (The University of Sydney)

Running time:0.04 s
Environment:1 core @ 2.5 Ghz (Python + C/C++)

Method Description:
We develop a new set abstraction module named
Semantics-Augmented Set Abstraction (SASA) for
point-based 3D detectors. It could largely enhance
the feature learning by helping extracted point
representations better focus on meaningful
foreground regions.
Parameters:
\gamma=1.0
Latex Bibtex:
@article{chen2022sasa,
title={SASA: Semantics-Augmented Set Abstraction
for Point-based 3D Object Detection},
author={Chen, Chen and Chen, Zhe and Zhang, Jing
and Tao, Dacheng},
journal={arXiv preprint arXiv:2201.01976},
year={2022}
}

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) 96.01 % 95.35 % 92.53 %
Car (Orientation) 96.00 % 95.29 % 92.42 %
Car (3D Detection) 88.76 % 82.16 % 77.16 %
Car (Bird's Eye View) 92.87 % 89.51 % 86.35 %
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