Method

Robust Augmentation R-FCN with ResNet-50 [at] [AR-FCN]


Submitted on 16 Jan. 2017 02:06 by
Yu Chen (CMU ECE)

Running time:0.19 s
Environment:GPU @ 2.5 Ghz (C/C++)

Method Description:
Utilize augmentation in recent objection detection
method R-FCN. And adding more training materials
like the night scene, the parking lot and background
to make the detection more robust.
Parameters:
NMS = 0.35, for better overlapped object.
\alpha=0.001
iteration times = 10000.
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) 81.24 % 75.49 % 66.00 %
Pedestrian (Detection) 53.16 % 43.88 % 35.58 %
Cyclist (Detection) 51.05 % 41.83 % 33.99 %
This table as LaTeX


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



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



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




eXTReMe Tracker