ANSIG An Analytic Signature for ANSIG An Analytic Signature for Permutation Invariant 2D Shape Representation José Jerónimo Moreira Rodrigues
ANSIG Outline Motivation: shape representation Permutation invariance: ANSIG Dealing with geometric transformations Experiments Conclusion Real-life demonstration
ANSIG MotivationANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Motivation The Permutation Problem
ANSIG MotivationANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Shape diversity
ANSIG MotivationANSIG Geometric transformations ExperimentsConclusion Real-life demonstration When the labels are known: Kendall’s shape ‘Shape’ is the geometrical information that remains when location/scale/rotation effects are removed. Limitation: points must have labels, i.e., vectors must be ordered, i.e., correspondences must be known
ANSIG MotivationANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Without labels: the permutation problem permutation matrix
ANSIG MotivationANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Our approach: seek permutation invariant representations
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration The analytic signature (ANSIG) of a shape
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Maximal invariance of ANSIG same signature equal shapes
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Maximal invariance of ANSIG Consider, such that Since, their first nth order derivatives are equal:
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Maximal invariance of ANSIG This set of equalities implies that - Newton’s identities The derivatives are the moments of the zeros of the polynomials
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Storing ANSIGs The ANSIG maps to an analytic function How to store an ANSIG?
Motivation ANSIG ANSIG Geometric transformations ExperimentsConclusion Real-life demonstration Storing ANSIGs 2) Approximated by uniform sampling: 1) Cauchy representation formula: 512
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Geometrictransformations
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG (Maximal) Invariance to translation and scale Remove mean and normalize scale:
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Sampling density
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Shape rotation: circular-shift of ANSIG Rotation
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Efficient computation of rotation Solution: maximum of correlation. Using FFTs, “time” domain frequency domain Optimization problem:
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Shape-based classification SHAPE TO CLASSIFY SHAPE 3 SHAPE 2 SHAPE 1 MÁXMÁX Similarity SHAPE2SHAPE2 DATABASE
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Experiments
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG MPEG7 database (216 shapes)
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Automatic trademark retrieval
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Robustness to model violation
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Object recognition
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Conclusion
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Summary and conclusion ANSIG: novel 2D-shape representation - Maximally invariant to permutation (and scale, translation) - Deals with rotations and very different number of points - Robust to noise and model violations Relevant for several applications Development of software packages for demonstration Publications: - IEEE CVPR IEEE ICIP Submitted to IEEE Transactions on PAMI
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Future developments Different sampling schemes More than one ANSIG per shape class Incomplete shapes, i.e., shape parts Analytic functions for 3D shape representation
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Real-lifedemonstration
ANSIG Motivation Geometric transformations ExperimentsConclusion Real-life demonstration ANSIG Shape-based image classfication Shape database Pre-processing: morphological filter operations, segmentation, etc. Image acquisition system Shape-based classification