Method

Domain Adaptive DPM [HA-SSVM]


Submitted on 22 Nov. 2013 10:34 by
Jiaolong Xu (Computer Vision Center (CVC))

Running time:21 s
Environment:1 core @ >3.5 Ghz (Matlab + C/C++)

Method Description:
We train a DPM (version 5) model with
virtual-world examples and we adapt it using
very few examples from the KITTI training set. We apply our Hierarchical Adaptive Structural SVM (HA-SSVM) to the DPM for the adaptation between the virtual-world examples and the real-world ones.
Parameters:
We have used 2000 virtual-world pedestrians and
2000 pedestrian free virtual-world images to train
a source domain model, and we adapted it to the
KITTI dataset by using randomly selected 200
pedestrians (approx. 10% of the mandatory
pedestrians in Kitti training set) and 2000
pedestrian free images. We use default DPM
parameters except the number of the components and
the parts as we set them to 4 and 5 respectively.
Latex Bibtex:
@Article{Xu2016IJCV,
author = {Jiaolong Xu and Sebastian Ramos and
David Vázquez and Antonio M. López},
title = {{Hierarchical Adaptive Structural SVM for Domain Adaptation}},
journal = {IJCV},
doi = {0.1007/s11263-016-0885-6},
year = {2016}
}

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) 58.76 % 43.87 % 38.81 %
This table as LaTeX


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




eXTReMe Tracker