Active Appearance Models AAM Shape Appearance Determine the best model parameters to reconstruct the image (Matthews and Gross) 42
Active Appearance Models Parameters AAM 0.12 -0.34 Shape 6.78 -12.2 Appearance 0.01 Determine the best model parameters to reconstruct the image (Matthews and Gross) 42
Active Appearance Models Image Parameters AAM 0.12 -0.34 Shape 6.78 -12.2 Appearance 0.01 Determine the best model parameters to reconstruct the image (Matthews and Gross) 42
Active Appearance Models Image Parameters AAM 0.12 -0.34 Shape 6.78 -12.2 Appearance 0.01 Modeling Determine the best model parameters to reconstruct the image (Matthews and Gross) 42
Active Appearance Models Image Parameters AAM 0.12 -0.34 Shape 6.78 -12.2 Appearance 0.01 Fitting Modeling Determine the best model parameters to reconstruct the image (Matthews and Gross) 42
AAM - Fitting “Simultaneous” • Now that we have defined an AAM, how should do the fitting? • Turns out, we can still use the “simultaneous” extension of the LK algorithm, ∆ λ , ∆ p || T ( z ) + A ∆ λ − I ( W ( z ; p )) − J ∆ p || 2 arg min • where, T ∂ W ( z ; p ) T T J = ∂ I ( W ( z ; p ) ∂ W ( z ; p ) = [ s 1 , . . . , s K ] ∂ W ( z ; p ) ∂ p ∂ p 43
AAM - Fitting “Inverse Composition” • Even faster if we use the inverse composition (IC) extension as well as “projecting out” the appearance, ∆ p || T ( z ) + J ic ∆ p − I ( W ( z ; p )) || 2 arg min null ( A ) • where, T ∂ W ( z ; 0 ) T T J ic = ∂ T ( W ( z ; 0 ) ∂ W ( z ; 0 ) = [ s 1 , . . . , s K ] ∂ W ( z ; 0 ) ∂ p ∂ p 44
IC AAM Fitting • Runs at 230 Hz on a 3.2GHz PC Shape Input Overlaid Overlaid Rendered Model Model Instance Instance (Matthews and Gross) 45
IC AAM Fitting • Runs at 230 Hz on a 3.2GHz PC Shape Input Overlaid Overlaid Rendered Model Model Instance Instance (Matthews and Gross) 45
Why We Need High Speed? Re-initialize model multiple times if tracking fails and still track in real time (Matthews and Gross) 46
Why We Need High Speed? Re-initialize model multiple times if tracking fails and still track in real time (Matthews and Gross) 46
Not All Is Peachy Original model does not handle occlusion well (Matthews and Gross) 47
Not All Is Peachy Original model does not handle occlusion well (Matthews and Gross) 47
AAM - Fitting “Robust Error Function” • To handle occlusions we can employ the robust error function, arg min ∆ λ , ∆ p η ( T ( z ) + A ∆ λ − I ( W ( z ; p )) − J ∆ p ) • where, T ∂ W ( z ; p ) T T J = ∂ I ( W ( z ; p ) ∂ W ( z ; p ) = [ s 1 , . . . , s K ] ∂ W ( z ; p ) ∂ p ∂ p 48
AAMs with Occlusion Modeling (Matthews and Gross) 49
AAMs with Occlusion Modeling (Matthews and Gross) 49
Applications • User Interfaces: • Mouse replacement • Automotive: Windshield Displays, Smart Airbags • Face Recognition: • Pose Normalization • Lipreading/Audio-Visual Speech Recognition • Rendering and Animation: • Low-Bandwidth Video Conferencing • Audio-Visual Speech Synthesis (Matthews and Gross) 50
User Interfaces: Head Pose • Mouse replacement (Matthews and Gross) 51
User Interfaces: Head Pose • Mouse replacement (Matthews and Gross) 51
User Interfaces: Gaze Tracking Driver Camera Exterior View Camera (Matthews and Gross) 52
User Interfaces: Gaze Tracking Driver Camera Exterior View Camera (Matthews and Gross) 52
Face Recognition: Pose Normalization (Matthews and Gross) 53
Face Recognition: Pose Normalization (Matthews and Gross) 53
Animation Generation (Matthews and Gross) 54
Animation Generation (Matthews and Gross) 54
Audio-Visual Speech Synthesis Jingle Bells, Jingle Bells, Jingle All the Way, ... (Matthews and Gross) 55
Recommend
More recommend
