Active Appearance Models Revisited

1
Active Appearance Models Revisited

• Iain Matthews and Simon Baker
• The Robotics Institute
• Carnegie Mellon University
• Presented By Brendan Morris

2
Full Algorithm
3
Algorithm Evaluation

• Inverse Compositional Warp vs. Additive Update
• Analytical Derivation of Steepest Decent Images
  vs. Numerical Approximation
• Projecting Out Appearance Variation vs. Fitting
  Appearance Parameters

4
AAM Construction

• Person Specific AAM
• 30 Training Images
• Multi-Person AAM
• 5 People
• 110 Training Images

5
Person Specific AAM
6
Multi-Person AAM
7
Experiment SetupTest Data

• Person Specific AAM
• 300 Frames
• Multi-Person AAM
• 900 Frames (180 Frames/Person)

8
Experiment SetupGround Truth

• Hand Initialization and Re-Initialization
• Tracking Through Sequences
• Inverse Composition Algorithm With Analytically
  Computed Steepest Decent Images and Projected-Out
  Appearance Variation (The Final Algorithm)
• Apparently Similar Results From Original Video
  and Reconstructed AAM Overlay Video

9
Experiment SetupProcedure

• Run Algorithm on Large Number of Inputs, Evaluate
  Results, and Average
• 20 Trials
• Input
• 1 Image of Test Data
• Ground Truth Shape, Appearance, and Similarity
  Parameters

10
Experiment SetupProcedure 2

• Initialize Algorithm by Random Perturbation of
  Ground Truth Parameters
• Shape N(0, a?), ? Eigenvalue From PCA
• Similarity Gaussian Noise Disturb 2 Points to
  Solve for Parameters
• Appearance Mean Appearance

11
Evaluation Metrics

• Measure Convergence
• Average Rate of Convergence
• RMS Error of Mesh Point Location
• RMS Error lt 1 Pixel for Convergence
• Average Frequency of Convergence
• Number of Times Algorithm Converged
• Limited to 20 Iterations

12
Experiment 1Update Rule

• Inverse Compositional Warp vs. Additive Update
• Perturbing Shape Parameters
• Similarity Transform Hides Problems

13
Experiment 1Update Rule 2

• Inverse Compositional Warp vs. Additive Update
• Perturbing Similarity Parameters

14
Experiment 1Update Rule 3

• Inverse Compositional Warp vs. Additive Update
• Perturbing Shape and Similarity Parameters

15
Experiment 2Steepest Decent Images

• Analytical Derivation of Steepest Decent Images
  vs. Numerical Approximation

16
Experiment 3Appearance Variation

• Projecting Out Appearance Variation vs. Fitting
  Appearance Parameters

17
Computation Efficiency

• Dual 2.4 GHz P4
• Person Specific AAM 19,977 Pixels
• 3 Shape Parameters
• 4 Similarity Transform Parameters
• 9 Appearance Parameters

18
Computational Efficiency 2
19
Computational Efficiency 3

• Inverse Compositional Update Fast
• Order of Additive Update
• Steepest Decent Much Faster Analytically than
  Numerically
• Explicit Modeling is Slower Because of More
  Parameters to Solve Per Iteration
• Error Image Computation More Involved
• 230 Frames/s For C Implementation

20
Conclusion

• Algorithm Has Accuracy of Inefficient Gradient
  Decent Algorithms with Speed of Linear Additive
  Update
• Speed of Convergence
• Frequency of Convergence
• Computational Cost
• Only Applicable to Independent AAMs
• Separate Shape and Appearance Parameters

21
The Robotics InstituteCarnegie Mellon University

• http//www.ri.cmu.edu/
• Real Time AAM Fitting Algorithms
• http//www.ri.cmu.edu/projects/project_448.html