Method

Cross modal network for monocular 3D object detection [CMAN]


Submitted on 2 Feb. 2022 14:57 by
Yuanzhouhan Cao (Beijing Jiaotong University)

Running time:0.15 s
Environment:1 core @ 2.5 Ghz (Python)

Method Description:
We propose a cross modal attention network (CMAN)
for 3D object detection. Our CMAN is built upon the
self-attention module which learns attention map
from single modal data such as RGB images. However
for 3D object detection, depth is an important cue.
RGB data contains rich appearance information but
lacks structure information. Our CMAN learns
attention map from depth data to capture structure
correlations and embed the structure correlation
with appearance information.
Parameters:
n.a.
Latex Bibtex:
@article{CMAN2022,
title={CMAN: Leaning Global Structure Correlation
for
Monocular 3D Object Detection},
author={ Yuanzhouhan Cao, Hui Zhang, Yidong Li,
Chao Ren, Congyan Lang},
journal={IEEE Trans. Intell. Transport. Syst.},
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) 89.74 % 83.74 % 65.35 %
Car (Orientation) 89.43 % 81.96 % 63.74 %
Car (3D Detection) 17.77 % 11.87 % 9.16 %
Car (Bird's Eye View) 25.89 % 17.04 % 12.88 %
Pedestrian (Detection) 49.73 % 34.96 % 30.92 %
Pedestrian (Orientation) 40.27 % 28.16 % 24.82 %
Pedestrian (3D Detection) 4.62 % 3.41 % 2.87 %
Pedestrian (Bird's Eye View) 5.24 % 3.96 % 3.18 %
Cyclist (Detection) 58.12 % 38.36 % 31.79 %
Cyclist (Orientation) 42.58 % 27.63 % 23.14 %
Cyclist (3D Detection) 1.59 % 1.05 % 1.11 %
Cyclist (Bird's Eye View) 1.76 % 1.48 % 1.17 %
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



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



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