Deformable Object Modeling and Rapid inference

1
Deformable Object Modeling and Rapid inference

• Leo Zhu
• Department of Statistics
• University of California Los Angeles
• Dec. 2007
• Collaborators Alan Yuille, Yuanhao Chen and
  Chenxi Lin

2
Deformable Object Parsing

• Deformable Object Parsing recovering the poses
  (location, scale and orientation) of object
  parts.
• Difficulties
• Large shape and appearance variations.
• Cluttered Background
• Occlusion, lighting, etc.

3
Representation Of Deformable Object

• Idea From Flat MRF to Hierarchical Composition
  Model
• Summarization Principle make reasonable
  independence assumptions
• Short-range and Long-range relations at multiple
  scales

4
Hierarchical Composition Model

• Formulation
• Image d States
• Parameters w
• Image Features is defined between the leaf nodes
  and image pixel.
• Horizontal Shape Priors at multiple levels
• Vertical constraints

5
Triplet Shape Descriptor

• Gaussian function on triplet shape descriptor
  (scale and rotation invariant).
• Shape deformations are modeled at multiple
  levels.

6
Bottom-Up and Top-Down Inference

• From the Bottom level to the Top Level
• Bottom-Up
• Hierarchical Composition
• Surround Suppression Independence Assumption
• Top-Down Change Proposals

7
Bottom-Up Inference

• From the Bottom Level to the Top Level
• Composition
• Pruning
• Surround Suppression
• Complexity empirically linear in the size of
  image and ranges of scale and orientation.

8
Bottom-up Inference

• Hierarchy

• Parsed Instances

Step 1 Composition

• Current Model

9
Bottom-up Inference

• Hierarchy

• Parsed Instances

Step 2 Pruning

• Current Model

10
Bottom-up Inference

• Hierarchy

• Parsed Instances

Step 3 Surround Suppression

• Current Model

11
Bottom-up Inference

• Hierarchy

• Parsed Instances

Summarization

• Current Model

12
Bottom-up Inference

• Hierarchy

• Parsed Instances

• Current Model

13
Bottom-up Inference

• Hierarchy

• Parsed Instances

• Current Model

14
Bottom-up Inference

• Hierarchy

• Parsed Instances

• Current Model

15
Bottom-up Inference

• Hierarchy

• Parsed Instances

• Current Model

16
Analysis I Computational Complexity
17
Analysis II Multi-level Performance
18
Summary

• Hierarchical Composition Model
• Rapid Inference/Parsing