Method

BirdNet: a 3D Object Detection Framework from LiDAR information [la] [BirdNet]


Submitted on 13 Feb. 2020 10:46 by
Carlos Guindel (Universidad Carlos III de Madrid)

Running time:0.11 s
Environment:Titan Xp (Caffe)

Method Description:
RESULTS UPDATED ON FEB 2020: As stated in
BirdNet+ (https://arxiv.org/abs/2003.04188), we
have found and fixed some bugs in our procedure
related to the transformation of the inference
results into camera coordinates. The method
remains unchanged from the one described in our
ITSC paper. (We acknowledge Alejandro Barrera for
the identification of those bugs.)

We present a LiDAR-based 3D object detection
pipeline entailing three stages. First, laser
information is projected into a novel cell
encoding for bird’s eye view projection. Later,
both object location on the plane and its heading
are estimated through a convolutional neural
network originally designed for image processing.
Finally, 3D oriented detections are computed in a
post-processing phase.
Parameters:
Resolution: 0.05 m
16 angle bins, without pool4, with ground.
Latex Bibtex:
@inproceedings{BirdNet2018,
author={J. Beltrán and C. Guindel and F. M.
Moreno and D. Cruzado and F. García and A. De La
Escalera},
booktitle={2018 21st International Conference on
Intelligent Transportation Systems (ITSC)},
title={BirdNet: A 3D Object Detection Framework
from LiDAR Information},
year={2018},
pages={3517-3523},
doi={10.1109/ITSC.2018.8569311},
ISSN={2153-0017},
month={Nov},}

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) 79.30 % 57.12 % 55.16 %
Car (Orientation) 79.20 % 56.94 % 54.88 %
Car (3D Detection) 40.99 % 27.26 % 25.32 %
Car (Bird's Eye View) 84.17 % 59.83 % 57.35 %
Pedestrian (Detection) 36.82 % 30.07 % 28.40 %
Pedestrian (Orientation) 27.12 % 21.83 % 20.56 %
Pedestrian (3D Detection) 22.04 % 17.08 % 15.82 %
Pedestrian (Bird's Eye View) 28.20 % 23.06 % 21.65 %
Cyclist (Detection) 64.91 % 47.64 % 44.59 %
Cyclist (Orientation) 62.69 % 45.03 % 41.88 %
Cyclist (3D Detection) 43.98 % 30.25 % 27.21 %
Cyclist (Bird's Eye View) 58.64 % 41.56 % 36.94 %
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