Method

DEVIANT: Depth EquiVarIAnt NeTwork for Moncoular 3D Object Detection [DEVIANT]
https://github.com/abhi1kumar/DEVIANT

Submitted on 18 Jul. 2022 02:09 by
Abhinav Kumar (Michigan State University (MSU))

Running time:0.04 s
Environment:1 GPU (Python)

Method Description:
Modern neural networks use building blocks such as convolutions that are equivariant to arbitrary 2D translations. However, these vanilla blocks are not equivariant to arbitrary 3D translations in the projective manifold. Even then, all monocular 3D detectors use vanilla blocks to obtain the 3D coordinates, a task for which the vanilla blocks
are not designed for. This paper takes the first step towards convolutions equivariant to arbitrary 3D translations in the projective manifold. Since
the depth is the hardest to estimate for monocular detection, this paper proposes Depth EquiVarIAnt NeTwork (DEVIANT) built with existing scale equivariant steerable blocks.
Parameters:
See the paper
Latex Bibtex:
@inproceedings{kumar2022deviant,
title={DEVIANT: Depth EquiVarIAnt NeTwork for Monocular 3D Object Detection},
author={Kumar, Abhinav and Brazil, Garrick and Corona, Enrique and Parchami, Armin and Liu, Xiaoming},
booktitle={European Conference on Computer Vision (ECCV)},
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) 94.42 % 86.64 % 76.69 %
Car (Orientation) 94.01 % 85.97 % 75.84 %
Car (3D Detection) 21.88 % 14.46 % 11.89 %
Car (Bird's Eye View) 29.65 % 20.44 % 17.43 %
Pedestrian (Detection) 74.27 % 55.16 % 50.21 %
Pedestrian (Orientation) 68.78 % 50.66 % 45.89 %
Pedestrian (3D Detection) 13.43 % 8.65 % 7.69 %
Pedestrian (Bird's Eye View) 14.49 % 9.77 % 8.28 %
Cyclist (Detection) 67.71 % 46.42 % 39.44 %
Cyclist (Orientation) 57.64 % 38.46 % 32.76 %
Cyclist (3D Detection) 5.05 % 3.13 % 2.59 %
Cyclist (Bird's Eye View) 6.42 % 3.97 % 3.51 %
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