Fast object detection

The ultimate goal of computer vision is to understand images. We describe methods to understand images at two levels. One is at the level of description of images which we produce using sentences. These sentences talk about the things that are present in the image and about where they are and what they are doing. Then we ask in what ways should we describe images. We introduce visual phrases that are composite chunks of meaning. We show that object detectors could be better at detecting some visual phrases than detecting single objects. This process of image understanding needs to use a lot of detectors. Running conventional object detectors at the rate required for image understanding could be very slow. We study fast object detection from an engineering perspective. We argue that a desirable object detector must: (1) be able to work with legacy templates; (2) be random access; (3) be able to trade accuracy versus speed; (4) have any-time property. We describe a method to have all of these features together for a fast detector. We apply these techniques to deformable parts model object detectors and show two orders of magnitude speed-up while adding their desirable features. We finally investigate the consequences of this architecture with a view of improving convolutional neural networks.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Computer Science

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2016

Author dc:creator

• Sadeghi, Mohammad Amin

Contributors dc:contributor

• Forsyth, David A
• Hoiem, Derek
• Ramanan, Deva
• Lazebnik, Svetlana
• Golparvar-fard, Mani

• Fast Object Detection
• Accurate Object Detection
• Visual Phrases
• Sentence Generation
• Vector Quantization.

Statement dc:rights

• Copyright 2015 Mohammad Amin Sadeghi

Language dc:language

en

Handle dc:identifier

http://hdl.handle.net/2142/89047

OAI identifier oai:identifier

oai:www.ideals.illinois.edu:2142/89047