Object Recognition as a Web Service

Miha Pleško (2014) Object Recognition as a Web Service. EngD thesis.

Preview

PDF Download (5Mb)

Abstract

The main focus of this work is on object categorization as a web service. A concrete system is studied, ViCoS Eye, which was developed by Visual Cognitive Systems Laboratory at Faculty of Computer and information Sciences. First we introduce the core computer vision algorithms that ViCoS Eye is based on. Support Vector Machine, Learnt-hierarhy-of-parts, Histogram of compositions descriptor and a combination of the last two are presented. Then we discuss open-source framework Apache Hadoop, that is used to efficiently run machine learning part of system. We also discuss another open-source framework, Storm, that is used for running categorization part of system. In addition, the efficient integration of algorithms onto the two frameworks is described. We split learning algorithm into smaller steps to follow the MapReduce paradigm. Similarly, prediction algorithm is split into smaller steps that are perfectly compatible with Storm framework. Finally we introduce, explain, implement and experimentally evaluate our own improvement of the system. Since we noticed that HoC descriptor is very sensitive when it comes to rotation of the input picture, we introduce an affine derivatives-based normalisation algorithm. Tests on standard Caltech-101 database were performed in two different contexts. The algorithm improves the results to some degree, when the input images are rotated. But a significant drop of classification accuracy was detected when the input objects were not deformed.

Actions (login required)