Method

ECP Faster R-CNN [at] [ECP Faster R-CNN]


Submitted on 19 Oct. 2018 12:52 by
Markus Braun (Daimler AG)

Running time:0.25 s
Environment:GPU @ 2.5 Ghz (Python)

Method Description:
We carefully adapted Faster R-CNN for the task of pedestrian detection. The VGG-16 model has been pre-trained on our EuroCity Persons dataset collected on-board a moving vehicle in 31 cities of 12 European countries. With over 238200 person instances manually labeled in over 47300 images, EuroCity Persons is nearly one order of magnitude larger than person datasets used previously for benchmarking.
Parameters:
VGG-16 used as feature extractor
Latex Bibtex:
@article{DBLP:journals/corr/abs-1805-07193,
author = {Markus Braun and
Sebastian Krebs and
Fabian Flohr and
Dariu M. Gavrila},
title = {The EuroCity Persons Dataset: {A} Novel Benchmark for Object Detection},
journal = {CoRR},
volume = {abs/1805.07193},
year = {2018},
url = {http://arxiv.org/abs/1805.07193},
archivePrefix = {arXiv},
eprint = {1805.07193},
timestamp = {Mon, 13 Aug 2018 16:46:52 +0200},
biburl = {https://dblp.org/rec/bib/journals/corr/abs-1805-07193},
bibsource = {dblp computer science bibliography, https://dblp.org}
}

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) 85.96 % 76.25 % 70.55 %
This table as LaTeX

2D object detection results.
This figure as: png eps pdf txt gnuplot




eXTReMe Tracker