Method

Joint Object Detection and Viewpoint Estimation using CNN features [FRCNN+Or]
https://github.com/cguindel/lsi-faster-rcnn

Submitted on 7 Jun. 2017 10:58 by
Carlos Guindel (Universidad Carlos III de Madrid)

Running time:0.1 s
Environment:GPU @ 1.5 Ghz (Python + C/C++)

Method Description:
We enhance the well-known Faster R-CNN method by adding viewpoint inference capabilities. Convolutional features are computed and shared for use in the tasks of region proposal, classification, and orientation estimation, for efficiency reasons.
Parameters:
Scale=500, 300 RPN proposals, 3 anchor scales and 3 anchor ratios, \N_b=8, \l_r=0.001, 150k iterations. Running time given for a NVIDIA Titan X Pascal.
Latex Bibtex:
@inproceedings{GuindelICVES,
author = {Guindel, Carlos and Martin, David and Armingol, Jose M.},
booktitle = {IEEE International Conference on Vehicular Electronics and Safety (ICVES)},
title = {{Joint Object Detection and Viewpoint Estimation using CNN features}},
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
Car (Detection) 89.60 % 78.59 % 68.69 %
Car (Orientation) 88.93 % 77.80 % 67.87 %
Pedestrian (Detection) 72.21 % 56.99 % 53.72 %
Pedestrian (Orientation) 67.92 % 52.96 % 49.61 %
Cyclist (Detection) 68.81 % 55.80 % 50.52 %
Cyclist (Orientation) 64.90 % 51.47 % 46.48 %
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



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



Orientation estimation 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




eXTReMe Tracker