Download presentation
Presentation is loading. Please wait.
Published byDelphia Chapman Modified over 8 years ago
1
Strong Supervision From Weak Annotation Interactive Training of Deformable Part Models ICCV 2011 2012/05/23
2
Abstract Provide a framework for large scale learning and annotation of structured models 1.Interactive labeling 2.Online learning
3
Interactive Part Localization: model and notation Define score of a part configuration: part configuration, image x, part tree model T = (V, E) location (x p, y p ), scale s p, orientation r p, aspect v p : value from part detector (HOG feature) : quadratic function on the relative displacement between parts Maximum likelihood solution:
4
Interactive Part Localization: incorporating user input score incorporating user input sequence of user input operations up to time step t. j(t): index of the part annotated by the user in time step t is the user’s label of the part location unary potential Simple extensions include allowing imperfect user responses
5
Interactive Part Localization: Dynamic Programming(1) The goal is to compute cache tables which are indexable by pixel location :an array storing the maximum likelihood solutions for part q at time step t ※ p->q the optimal solution conditioned on placing p at position :
6
Interactive Part Localization: Dynamic Programming(2) : distance transform operation : cost associated with part q being at an offset of from p
7
Interactive Part Localization: Dynamic Programming(3) Using the pre-computed cache table
8
Interactive Part Localization: Algorithm
9
Online Structured Learning maximum margin structured learning problem (structured SVM): search optimal vector of weights w* error function: vector of HOG features + squared distances between adjacent parts x i image, y i = ɵ, penalty stochastic gradient descent (SGD): gradient:
10
Online Structured Learning: Algorithm
11
Experiments
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.