Object recognition and localisation - Diplomarbeit · Object recognition and localisation Melanie...
Transcript of Object recognition and localisation - Diplomarbeit · Object recognition and localisation Melanie...
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Object recognition and localisationDiplomarbeit
Melanie Dietz
TU Munchen
November 24, 2005
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
1 Neural NetworksBasicsFeed forward vs. backward
2 Vision and localisationThe human visionHierarchical NetworksHMAX
3 Diploma thesisThe hierarchical networkAdding feedback connectionsTraining and analysis
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Basics
Modelled on biological system
It works! (look in a mirror)
Components:1 nodes2 weighted connections
many types (unidirectional, bidirectional, layered,hierarchical, ...)
empirical design or training of weights
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Feed forward vs. backward
Feed forward’A feed-forward network can be viewed as a graphicalrepresentation of parametric function which takes a set ofinput values and maps them to a corresponding set ofoutput values (Bishop, 1995).’⇒: classification and prediction applications
Feed backward
No declared output nodesInstead output is fed back as input signal
⇒: optimization applications
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
The visual cortex
The retina
LGN
Visual cortical areas
Visual pathways
Hierarchical structure
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Hierarchical Networks
nodes are arranged in layers
information gain per layer increases
number of nodes per layer decreases
new concepts:
physical position in lower layeresreceptive fieldstranslation invariance (2D)
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Simple and complex cells
Hubel and Wiesel (1962)
Simple cells
small receptive fieldsstrong phase dependence
Complex cells
larger receptive fieldsno phase dependence
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
HMAX
Riesenhuber and Poggio(1982)
pooling operation→ view invariance(3D)
simple and complexcells
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
The hierarchical network
implementation modelled on HMAX
First layer: noise reduction
Second layer: S1 (orientation filter)C1 (orientation collection)
Third layer: S2 (simple features)C2 (complex features)
Forth layer: view-tuned cells
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Orientation detection
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Training and analysis
usage of BP for VTU’s
creation of trainig data
composition of training sets
comparison to HMAX (Matlab version)
Objectrecognition
andlocalisation
Melanie Dietz
Outline
NeuralNetworks
Basics
Feed forward vs.backward
Vision andlocalisation
The humanvision
HierarchicalNetworks
HMAX
Diploma thesis
The hierarchicalnetwork
Adding feedbackconnections
Training andanalysis
Adding feedback connections
not only ventral stream but also dorsal stream
backward connections for asking where the object was
usage of shape AND color for object recognition