Method

Multi-Receptive Field Pillars [MuRF]


Submitted on 21 Jan. 2020 13:27 by
liu ziyi (University at Buffalo)

Running time:0.05 s
Environment:GPU @ 1.5 Ghz (Python + C/C++)

Method Description:
We propose a cloud point based 3D object detection
framework that accounts for both contextual and
local information by leveraging multi-receptive
field pillars.
Parameters:
\voxel_size=0.16, \receptive_size=0.6,0.16,0.32
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) 95.74 % 92.74 % 87.64 %
Car (Orientation) 0.63 % 1.75 % 2.14 %
Car (3D Detection) 84.81 % 75.11 % 69.99 %
Car (Bird's Eye View) 91.57 % 88.56 % 83.46 %
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