lec23

lec23 - Lecture 23: November 10, 10 HW5 due today; hw6...

Info iconThis preview shows pages 1–9. Sign up to view the full content.

View Full Document Right Arrow Icon
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 1 Lecture 23: November 10, 10 • HW5 due today; hw6 posted today • Exam 2, Dec 1, during class time, closed book, closed notes – Similar in style to Exam 1 – Content: Material from Lec 12 onwards, detailed list will be given later. •R e v i e w – Support vector machine – HOG-based pedestrian detector (Dalal-Triggs) •T o d a y – Part-based detection – “Bag of Words” approach
Background image of page 1

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 2 Part-based Object Recognition • Global features are sensitive to occlusions • Local features not affected by occlusions in other parts but the set of features may be affected severely • Part-based analysis offers one way to overcome effects of occlusion • Objects may be considered to be made of parts, e.g. a human body has a head, torso, legs and arms – Hierarchical, parts may also be made of smaller parts • Object model consists of parts (their shapes or a classifier for them) and relation between the parts – e.g. head is not to the side of the torso – How to find the parts and how to verify the relations?
Background image of page 2
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 3 Human (pedestrian) Detection • Objective: Detect multiple, possibly partially occluded, standing/walking humans from static images in cluttered environments • Differences from face detection problem – No consistent region gray level pattern (varied clothing) – Frequent inter-object occlusion • Describe a method presented in Wu-Nevatia, ICCV 2005
Background image of page 3

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 4 Key Techniques • Part based human representation • Edgelet features – Other dense feature types, e.g. HOG could also be used • Decisions by computing joint likelihoods of multiple objects – The projections of objects are not independent of each other – Inter-object occlusion reasoning goes here
Background image of page 4
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 5 Basic Assumptions • Humans are walking or standing on a ground plane • Viewed from frontal view point (relaxed in newer work) • Camera looks down towards the ground • Average heights ground plane 3D scene 2D image y x 45 o
Background image of page 5

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 6 Part Representation Full-body Head- shoulder Torso Legs h w 0.3 h 0.27 h 0.48 h 0.5 h
Background image of page 6
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 7 Edgelet Features • Edgelet: small segment of edge template • Match with edge intensity map (magnitude of Sobel gradient) • Overcomplete: 857,604 overall edgelet features 0 0 1 2 3 4 5 1 23 4 5 Orientation quantization Sobel convolution result line arc symmetric pair
Background image of page 7

Info iconThis preview has intentionally blurred sections. Sign up to view the full version.

View Full DocumentRight Arrow Icon
USC CS574: Computer Vision, Fall 2010 Copyright 2010, by R. Nevatia 8 Learning Part Detectors • Boosted cascade-like classifier using edgelet features
Background image of page 8
Image of page 9
This is the end of the preview. Sign up to access the rest of the document.

This note was uploaded on 11/23/2010 for the course CS 574 taught by Professor Ramnevatia during the Fall '10 term at USC.

Page1 / 36

lec23 - Lecture 23: November 10, 10 HW5 due today; hw6...

This preview shows document pages 1 - 9. Sign up to view the full document.

View Full Document Right Arrow Icon
Ask a homework question - tutors are online