Extending Pictorial Structures for Object Recognition - PowerPoint PPT Presentation

1 / 1
About This Presentation
Title:

Extending Pictorial Structures for Object Recognition

Description:

Aim: Recognizing object categories (e.g. cows, not Daisy') Challenges : ... GMMk is the Gaussian mixture model over the intensity values of the kth ... – PowerPoint PPT presentation

Number of Views:64
Avg rating:3.0/5.0
Slides: 2
Provided by: aiSta
Category:

less

Transcript and Presenter's Notes

Title: Extending Pictorial Structures for Object Recognition


1
Extending Pictorial Structures for Object
Recognition
Aim Recognizing object categories (e.g. cows,
not Daisy)
  • Outline (z1) minimum truncated chamfer distance
    over multiple exemplars.

Results
Outline of the parts detected are shown overlaid
on the images. Correct detection is when the
object is found at the right position.
Challenges
2. Texture (z2) maxk log(Pr(DiGMMk))
  • Di is the region defined by the ith part.
  • GMMk is the Gaussian mixture model over the
    intensity values of the kth segmented training
    image.
  • Likelihood of (z1,z2) is modelled as a 2D normal
    distribution.
  • Intra-class variability in the shape, texture
    and spatial arrangement of parts.
  • Partial or complete occlusion of parts.
  • Slight variations in the pose of the object.
  • Changes in appearance due to lighting conditions.
  • 3. Configuration
  • Configuration forms a complete graph.
  • ?(li,lj) const, if valid configuration
  • 0, otherwise
  • Valid configurations are those seen in training
    video sequences.

Valid positions for the centroid of the half limb.
Method Pictorial Structures (PS)
Model Components
1. Outline describes the shape of a part. 2.
Texture describes the intensity values of points
inside the region defined by the outline. 3.
Configuration describes the spatial
relationships between certain pairs of parts.
Outline
Learning model parameters
  • Learning layered pictorial structures ICVGIP
    04

(x,y)
Tree-cascade of classifiers. Leaf nodes
contain scaled and rotated versions of exemplars.
2D Parts Configuration
Texture
Tree Structure
The connections between the forelegs results
in better matching of the torso and legs.
Vs.
  • 20 videos of 45 frames each.

Complete Graph
  • Parameters of likelihood functions are learnt by
    computing
  • (z1,z2) for ve and ve examples.

ROC Curves
Recognition
500 cow images 5000 ve examples
  • MAP estimation of poses of all parts using loopy
    belief propagation leading to improvements over
    previous methods.
  • Affine refinement of parts to minimize chamfer
    distance.

Extensions over PS
Write a Comment
User Comments (0)
About PowerShow.com