@mastersthesis{Kloss2018,
title = {Boosted Projections and Low Cost Transfer Learning Applied to Smart Surveillance},
author = {Ricardo Barbosa Kloss},
url = {https://www.dropbox.com/s/dpkkv16zkardxdx/main.pdf?dl=0},
year = {2018},
date = {2018-02-01},
school = {Federal University of Minas Gerais},
abstract = {Computer vision is an important area related to understanding the world through images.
It can be used as biometry, by verifying if a given face is of a certain identity, used to look for crime perpetrators in an airport blacklist, used in human-machine interactions and other goals.
Since 2012, deep learning methods have become ubiquitous in computer vision achieving breakthroughs, and making possible for machines, for instance, to perform face verification with human-level skill.
This work tackles two computer vision problems and is divided in two parts.
In one we explore deep learning methods in the task of face verification and in the other the task of dimensionality reduction.
Both tasks have large importance in the fields of machine learning and computer vision.
We focus on their application in smart surveillance.
Dimensionality reduction helps alleviate problems which usually suffer from a very high dimensionality, which can make it hard to learn classifiers.
This work presents a novel method for tackling this problem, referred to as Boosted
Projection.
It relies on the use of several projection models based on
Principal Component Analysis or Partial Least Squares to build a more compact and richer data representation.
Our experimental results demonstrate that the proposed approach outperforms many baselines and provides better results when compared to the original dimensionality reduction techniques of partial least squares.
In the second part of this work, regarding face verification, we explored a simple and cheap technique to extract deep features and reuse a pre-learned model.
The technique is a transfer learning that involves no fine-tuning of the model to the new domain.
Namely, we explore the correlation of depth and scale in deep models, and look for the layer/scale that yields the best results for the new domain, we also explore metrics for the verification task, using locally connected convolutions to learn distance metrics.
Our face verification experiments use a model pre-trained in face identification and adapt it to the face verification task with different data, but still on the face domain.
We achieve 96.65% mean accuracy on the Labeled Faces in the Wild dataset and 93.12% mean accuracy on the Youtube Faces dataset which are in the state-of-the-art.},
keywords = {Computer vision, Deep Learning, Forensics, Machine Learning, Surveillance},
pubstate = {published},
tppubtype = {mastersthesis}
}