CSEIT411830 | Published - 25 April 2018 | March-April-2018 [ (4 ) 1 : 178-182 ]
National Conference on Recent Advances in Computer Science and IT (NCRACIT)
International Journal of Scientific Research in Computer Science, Engineering and Information Technology
© 2018 IJSRCSEIT | Volume 4 | Issue 1 | ISSN : 2456-3307
178
Subspace-Based Adaptation of Detectors for Video
Sangeeta Bhan
1
, Sunil Dalal
2
, Pankaj Choudhary
3
1
Computer Science and Engineering, Delhi Technological University, Delhi, India
23
Department of Information Technology BGSB University, Rajouri, J&K, India
sangeetabhan7@gmail.com
1
, sunildalal57@gmail.com
2
, pschoudhary@bgsbu.ac.in
3
ABSTRACT
Object detection in videos has always been a challenging problem to work with. Detection of a particular
class object plays an important role in many real-world applications. Since the domain of source and target
video vary significantly, classifier being trained on source video does not give expected results on the target
video. Thus, domain adaptation techniques are used, one of which is Subspace Based Adaptation. In this
technique, first, we compute both source and target subspace from the features collected. Since we do not
have target data directly, we use different ways to get data from the target video. Compute subspace after
collecting the data from both source and target videos. Eigen vectors describe this generated source and target
subspaces.
Keywords: Subspace, Detector, Adaptation, Histogram.
I. INTRODUCTION
In detection the object of the particular class, it is
typically presumed that source and target data have
the same distribution. However, it is not true in real
world applications. We are training our simple
detector model from the source video with the help
of its annotation file. Initially, the problem is to
collect the negatively labeled data from video. We
have used Hard Negative Mining to collect this
negatively labeled data. Next, we extract the features
of these data using Histogram of Oriented Gradients.
With the help of Background Subtraction using MoG,
data samples were collected from the target video for
subspace calculation. PCA is used to find the best
Eigen vectors for both source and target dataset.
Once we have calculated the subspaces using PCA,
transforms the source subspace coordinate system
into the target aligned source subspace coordinate
system by aligning the source basis vectors with the
target ones. Generate training data by mapping
source data into this target aligned source subspace.
Now, we train our linear SVM model using this new
labeled data. Iteratively use online samples or
bounding boxes detected on target video to generate
new detector model.
II. RELATED WORK AND BACKGROUND
THEORY
Over the past several years, many different have
been pro-posed to detect objects of a particular class
in videos and images. The main problem in object
detection is varying domains of source and target
data. So the main issue is to find out the relationship
between these two domains. Domain adaptation is a
widely used technique in computer vision and
language processing. A classical strategy related to
our work consists of learning a new domain-
invariant feature representation by a new projection