Challenges in Computer Vision

builderanthologyΤεχνίτη Νοημοσύνη και Ρομποτική

19 Οκτ 2013 (πριν από 3 χρόνια και 7 μήνες)

268 εμφανίσεις

Challenges in Computer Vision

Understanding the ”seeing machine”


The input (images)


The output (shapes, actions?, diagnosis?)


The mapping (statistics, learning)


The computation (algorithms)

Images

Shapes

Statistics

Computation

Fundamental problems


Input space is high (infinite) dimensional


Modeling of shape


Mapping is nonlinear


Generalization from few examples


Finite memory and computational time


Fundamental challenge

Computational efficient, statistical optimal mapping from
images to models/actions:


The optimal mapping is unreachable (Kolmogorov)

Infinite computation time using optimal mapping (Ryabko)

So only hacking is left, so let’s hack

Only theories that proove themselves in practice


are good theories

Desired properties


Generalisation
-
> metric in input and model space


Universality
-
> Flexible/scalable models


Fast convergence
-
> ”least committed priors” and
”good learning”

Concrete challeges


Statistical well
-
founded metric on images


Geometrical metrics on shapes


Universal shape models


Information reduction or ”visual attention”


Marginalisation over hidden model parameters


Computational efficient approximative methods


Statistics on trees/graphs

Metrics on images


Scale
-
space


Independent component analysis


Geometry of images

Metrics on shapes


Shape
-
space theory


Grenander’s ”Theory of shape”


Invariant parametrisations


Brownian motions


Warps of embedding spaces


Lie
-
group methods

Universal shape models


Fourier descriptors


Landmark representations


Medial models


Level sets

Information reduction


Dimensionality reduction


Feature selection


”AdaBoost”

Marginalisation over hidden parameters


Common sense


Mean field analysis?

Computational efficient methods


PDEs


Particle filters


Hierachical representations


Sequential testing