Method

AVOD-FPN + single SHPL [SparsePool]
https://github.com/ZiningWang/Sparse_Pooling

Submitted on 18 Nov. 2019 06:01 by
Zining Wang (University of California, Berkeley)

Running time:0.13 s
Environment:8 cores @ 2.5 Ghz (Python)

Method Description:
AVOD-FPN + single directional sparse non-
homogeneous pooling (cam->LIDAR). Feature layers
are fused right before region proposal.
Parameters:
Historical results (Jan 2018) (Hard to reproduce
claimed results on ped and cyc from available
methods)
Latex Bibtex:
@article{wang2017fusing,
title={Fusing bird view lidar point cloud and
front view camera image for deep object
detection},
author={Wang, Zining and Zhan, Wei and
Tomizuka, Masayoshi},
journal={arXiv preprint arXiv:1711.06703},
year={2017}
}

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
Pedestrian (Detection) 50.94 % 39.43 % 35.78 %
Pedestrian (Orientation) 43.86 % 33.35 % 29.99 %
Pedestrian (3D Detection) 37.84 % 30.38 % 26.94 %
Pedestrian (Bird's Eye View) 43.33 % 34.15 % 31.78 %
Cyclist (Detection) 44.21 % 36.26 % 32.57 %
Cyclist (Orientation) 43.33 % 34.56 % 31.09 %
Cyclist (3D Detection) 40.87 % 32.61 % 29.05 %
Cyclist (Bird's Eye View) 43.55 % 35.24 % 30.15 %
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




eXTReMe Tracker