learning based contour detection contour based object
play

Learning-based Contour Detection & Contour-based Object Detection - PowerPoint PPT Presentation

1 Learning-based Contour Detection &


  1. ����������������������������������������������������������������� 1 Learning-based Contour Detection & Contour-based Object Detection Iasonas Kokkinos Department of Applied Mathematics Galen Group Ecole Centrale de Paris INRIA-Saclay 21 January, 2011 Visual Geometry Group, Oxford

  2. ����������������������������������������������������������������� 2 Talk outline Boundary Detection (35’) Logistic regression and Anyboost F-measure Boosting MIL and boundary detection Monte Carlo approximations for large-scale datasets Monte Carlo approximations for large-scale datasets Appearance descriptors and boundary detection Object Detection (15’) Coarse-to-fine inference (parsing) Model learning

  3. ����������������������������������������������������������������� 3 Image Contours Object/Surface Boundaries (edges)

  4. ����������������������������������������������������������������� 4 Image Contours Symmetry axes (ridges/valleys)

  5. ����������������������������������������������������������������� 5 A biref anlaogy wtih txet Waht mttares is waht hppaens on wrod bandouries Mocpera iwht htsi (compare with this) Concrete evidence that our visual system employs boundary detection Contour-based approaches: shape matching, segmentation, recognition,..

  6. ����������������������������������������������������������������� 6 How can we detect boundaries? Filtering approaches Canny (1984), Morrone and Owens (1987), Perona and Malik (1991),.. Scale-Space approaches Witkin, A. P. "Scale-space filtering", IJCAI (1983) Tony Lindeberg `Edge Detection and Ridge Detection with Automatic Scale Selection.’, IJCV, 30(2), 117-156, (1998) Variational approaches M. Kass, A. Witkin and D. Terzopoulos, `Snakes: Active Contour Models’, ICCV (1987) V. Caselles, R. Kimmel, G. Sapiro: Geodesic Active Contours. IJCV22(1): 61-79 (1997) K. Siddiqi, Y. Lauzière, A. Tannenbaum, S. Zucker: Area and length minimizing flows for shape segmentation. IEEE TIP 7(3): 433-443 (1998) Gestalt-based approaches Agnès Desolneux, Lionel Moisan, Jean-Michel Morel: Meaningful Alignments. International Journal of Computer Vision 40(1): 7-23 (2000)

  7. ����������������������������������������������������������������� 7 Learning-based approaches Boundary or non-boundary? Use human-annotated segmentations Use human-annotated segmentations S. Konishi, A.Yuille, J. Coughlan, S.C. Zhu, “Statistical Edge Detection: Learning and Evaluating Edge Cues”, IEEE PAMI, 2003 D. Martin, C. Fowlkes, J. Malik. "Learning to Detect Natural Image Boundaries Using Local Brightness, Color and Texture Cues", IEEE PAMI, 2004

  8. ����������������������������������������������������������������� 8 Progress during the last 40 years Humans Berkeley gPb, ‘08 Berkeley PB, ‘04 Canny+ Hysteresis Prewitt, 1965

  9. ����������������������������������������������������������������� 9 A closer look into gPb: features Local features (Pb, 2004) Global features (gPb, 2008) N-Cuts eigenvectors r θ (x,y) In specific:

  10. ����������������������������������������������������������������� 10 A closer look into gPb: classifier Logistic regression

  11. ����������������������������������������������������������������� 11 Talk outline Boundary Detection (35’) Logistic regression and Anyboost F-measure Boosting MIL and boundary detection Monte Carlo approximations for large-scale datasets Monte Carlo approximations for large-scale datasets Appearance descriptors and boundary detection Object Detection (15’) Coarse-to-fine inference (parsing) Model learning

  12. ����������������������������������������������������������������� 12 Learning Given: Training set of feature-label pairs Wanted: `simple’ that `works well’ on `simple’: quantified by VC dimension, curvature,… `works well’: quantified by loss criterion

  13. ����������������������������������������������������������������� 13 Logistic regression Linear function: Log-likelihood of training pair: Loss function: Optimization: Newton-Raphson (IRLS)

  14. ����������������������������������������������������������������� 14 Anyboost Additive form: At each round, add optimal pair See training cost as function of Steepest descent direction: Find `closest’ to Adaboost: exponential loss sign weight

  15. ����������������������������������������������������������������� 15 Side-by-side Logistic regression Anyboost � Additive � Linear � Summands: features � Summands: features � Summands: weak learners � Summands: weak learners � fixed � added `on the fly’ � � : Newton-Raphson : Coordinate descent � Cost: exponential loss (Adaboost) � Cost: minus label log likelihood Connections: M. Collins, R. Schapire, Y. Singer `Logistic Regression, AdaBoost and Bregman Distances’ COLT (2000)

  16. ����������������������������������������������������������������� 16 A compact combination Goal: quick classification, using small (e.g. ) feature set. � Additive � (linear part) � Remaining summands: weak learners (nonlinearities) � : Newton-Raphson, at each iteration � Slower, but off-line � Cost?

  17. ����������������������������������������������������������������� 17 Talk outline Boundary Detection (35’) Logistic regression and Boosting, Anyboost F-measure Boosting MIL and boundary detection Monte Carlo approximations for large-scale datasets Monte Carlo approximations for large-scale datasets Appearance descriptors and boundary detection Object Detection (15’) Coarse-to-fine inference (parsing) Model learning

  18. ����������������������������������������������������������������� 18 Cost function for training Training set Classifier Loss additive additive non-additive: F-measure, Area Under Curve (AUC),… - potentially better suited for the problem - but also potentially non-convex (local optimality) T. Joachims, `A Support Vector Method for Multivariate Performance Measures’, ICML, 2005 M. Jansche, `Maximum Expected F-Measure Training Of Logistic Regression Models’, EMNLP, 2005 M. Ranjbar, G. Mori and Y. Wang `Optimizing Complex Loss Functions in Structured Prediction’ ECCV, 2010

  19. ����������������������������������������������������������������� 19 F-measure Goal: deal with unbalanced datasets (many negative) no reward for true negative decisions Predicted label true positives misses false alarms precision recall F-measure: geometric mean of precision and recall

  20. ����������������������������������������������������������������� 20 F-measure approximation predicted label differentiable approximation approximate F-measure M. Jansche, ‘Maximum Expected F-Measure Training Of Logistic Regression Models’, EMNLP, 2005

Download Presentation
Download Policy: The content available on the website is offered to you 'AS IS' for your personal information and use only. It cannot be commercialized, licensed, or distributed on other websites without prior consent from the author. To download a presentation, simply click this link. If you encounter any difficulties during the download process, it's possible that the publisher has removed the file from their server.

Recommend


More recommend