A Three-Stage RGBD Architecture for Vehicle Detection Using Convolutional Neural Networks [st] [4D-MSCNN+CRL]

Submitted on 28 Dec. 2018 01:34 by
Pedro Augusto Pinho Ferraz (Pontifical Catholic University of Minas Gerais - PUC Minas)

Running time:0.2 s
Environment:GPU @ 2.5 Ghz (Matlab + C/C++)

Method Description:
We introduce a three-stage architecture that extracts depth information from stereo images and merges it with the traditional RGB image, giving rise to a four-channel RGBD input to improve detection of vehicles using convolutional neural networks. This submission uses the CRL algorithm to generate disparity maps and MSCNN as detection CNN. The details of architecture and implementation are presented in our paper.
Please check our paper and Github for more information.
Latex Bibtex:

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) 92.40 % 89.37 % 77.00 %
This table as LaTeX

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

eXTReMe Tracker