Representing images and regions for object recognition

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2015

Öğrenci: İLKER BUZCU

Danışman: ABDULLAH AYDIN ALATAN

Özet:

We can represent images in entirely different ways, in order to fulfill different purposes. For object recognition, power of a representation comes from its discriminative ability. In this thesis work, handcrafted representations that dominated the last decade of computer vision are evaluated against the current paradigm of Deep Learning, to try and pinpoint the reasons behind why and how the fairly old Artificial Neural Network (ANN) framework suddenly emerged as the state of the art in discriminative representations. We observe, through our experiments, that true capabilities of Deep ANN's can only be achieved by having very large amounts of labeled data that have been made available only recently. This thesis work also deals with ensembles of both handcrafted and ANN based approaches to reinforce the new technology with years of established knowledge behind handcrafted feature based approaches. For this purpose, we propose a novel extension, based on Fisher Vectors, to the well known Selective Search algorithm, called the Fisher-Selective Search algorithm, and obtain a 10% relative increase in Average Precision at virtually no additional computation cost.