Method

Faraway-Frustum: Dealing with Lidar Sparsity for 3D Object Detection using Fusion [la] [Faraway-Frustum]
https://github.com/dongfang-steven-yang/faraway-frustum

Submitted on 22 Dec. 2020 21:07 by
Haolin Zhang (CITR Lab, The Ohio State University)

Running time:0.1 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
we introduce faraway-frustum: a novel fusion
strategy for detecting faraway objects. The
main strategy is to depend solely on the 2D
vision for recognizing object class, as
object shape does not change drastically with
an increase in depth, and use pointcloud data
for object localization in the 3D space for
faraway objects. For closer objects, we use
learned pointcloud representations instead,
following state-of-the-art.
Parameters:
Distance threshold for faraway pedestrians:
60 meters
Distance threshold for faraway cars: 75
meters
Latex Bibtex:
@article{zhang2020faraway,
title={Faraway-Frustum: Dealing with Lidar
Sparsity for 3D Object Detection using
Fusion},
author={Zhang, Haolin and Yang, Dongfang
and Yurtsever, Ekim and Redmill, Keith A and
Ozguner, Umit},
journal={arXiv preprint arXiv:2011.01404},
year={2020}
}

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) 95.81 % 93.99 % 91.72 %
Car (3D Detection) 87.45 % 79.05 % 76.14 %
Car (Bird's Eye View) 91.90 % 88.08 % 85.35 %
Pedestrian (Detection) 67.88 % 57.35 % 54.42 %
Pedestrian (3D Detection) 46.33 % 38.58 % 35.71 %
Pedestrian (Bird's Eye View) 52.15 % 43.85 % 41.68 %
Cyclist (Detection) 85.43 % 73.63 % 66.64 %
Cyclist (3D Detection) 77.36 % 62.00 % 55.40 %
Cyclist (Bird's Eye View) 79.65 % 64.54 % 57.84 %
This table as LaTeX


2D object detection 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



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



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