Active Appearance ModelsOverviewWhat are we trying to do?Appearance ModelsFirst approach: Active Shape Model (ASM)First Approach: ASM (cont.)Lessons learnedCombined Appearance ModelsHow to generate a CAMHow to generate a CAM (cont.)How to generate a CAM (cont.)How to generate a CAM (cont.)CAM PropertiesCAM Properties (cont.)CAM Properties (cont.)CAM Properties (cont.)How to interpret unseen exampleHow to interpret unseen example (cont.)AAM: TrainingAAM: Training (cont.)AAM: Training (cont.)AAM: Training (cont.)AAM: SearchStarting approximationStarting approximation (cont.)AAM: Search (cont.)Experimental resultsExperimental results (cont.)Experimental results (cont.)Experimental results (cont.)Experimental results (cont.)Experimental results (cont.)Experimental results (cont.)Constrained AAMsConstrained AAMsConstrained AAMs (cont.)Constrained AAMsConclusionsConclusions (cont.)Active Appearance ModelsEdwards, Taylor, and CootesPresented by Bryan RussellOverview Overview of Appearance Models Combined Appearance Models Active Appearance Model Search Results Constrained Active Appearance ModelsWhat are we trying to do? Formulate model to “interpret” face images– Set of parameters to characterize identity, pose, expression, lighting, etc.– Want compact set of parameters – Want efficient and robust modelAppearance Models Eigenfaces (Turk and Pentland, 1991)– Not robust to shape changes– Not robust to changes in pose and expression Ezzat and Poggio approach (1996)– Synthesize new views of face from set of example views– Does not generalize to unseen facesFirst approach: Active Shape Model (ASM) Point Distribution ModelFirst Approach: ASM (cont.) Training: Apply PCA to labeled images New image– Project mean shape– Iteratively modify model points to fit local neighborhoodLessons learned ASM is relatively fast ASM too simplistic; not robust when new images are introduced May not converge to good solution Key insight: ASM does not incorporate all gray-level information in parametersCombined Appearance Models Combine shape and gray-level variation in single statistical appearance model Goals:– Model has better representational power– Model inherits appearance models benefits– Model has comparable performanceHow to generate a CAM Label training set with landmark points representing positions of key features Represent these landmarks as a vector x Perform PCA on these landmark vectorsHow to generate a CAM (cont.) We get: Warp each image so that each control point matches mean shape Sample gray-level information g Apply PCA to gray-level dataHow to generate a CAM (cont.) We get: Concatenate shape and gray-level parameters (from PCA) Apply a further PCA to the concatenated vectorsHow to generate a CAM (cont.) We get:CAM Properties Combines shape and gray-level variations in one model– No need for separate models Compared to separate models, in general, needs fewer parameters Uses all available informationCAM Properties (cont.) Inherits appearance model benefits– Able to represent any face within bounds of the training set– Robust interpretation Model parameters characterize facial featuresCAM Properties (cont.) Obtain parameters for inter and intra class variation (identity and residual parameters) – “explains” faceCAM Properties (cont.) Useful for tracking and identification– Refer to: G.J.Edwards, C.J.Taylor, T.F.Cootes. "Learning to Identify and Track Faces in Image Sequences“. Int. Conf. on Face and Gesture Recognition, p. 260-265, 1998. Note: shape and gray-level variations are correlatedHow to interpret unseen example Treat interpretation as an optimization problem– Minimize difference between the real face image and one synthesized by AAMHow to interpret unseen example (cont.) Appears to be difficult optimization problem (~80 parameters) Key insight: we solve a similar optimization problem for each new face image Incorporate a-priori knowledge for parameter adjustments into algorithmAAM: Training Offline: learn relationship between error and parameter adjustments Result: simple linear modelAAM: Training (cont.) Use multiple multivariate linear regression– Generate training set by perturbing model parameters for training images– Include small displacements in position, scale, and orientation– Record perturbation and image differenceAAM: Training (cont.) Important to consider frame of reference when computing image difference– Use shape-normalized representation (warping)– Calculate image difference using gray level vectors:AAM: Training (cont.) Updated linear relationship: Want a model that holds over large error range Experimentally, optimal perturbation around 0.5 standard deviations for each parameterAAM: Search Begin with reasonable starting approximation for face Want approximation to be fast and simple Perhaps Viola’s method can be applied hereStarting approximation Subsample model and image Use simple eigenface metric:Starting approximation (cont.) Typical starting approximations with this methodAAM: Search (cont.) Use trained parameter adjustment Parameter update equation:Experimental results Training:– 400 images, 112 landmark points– 80 CAM parameters– Parameters explain 98% observed variation Testing:– 80 previously unseen facesExperimental results (cont.) Search results after initial, 2, 5, and 12 iterationsExperimental results (cont.) Search convergence: – Gray-level sample error vs. number of iterationsExperimental results (cont.) More reconstructions:Experimental results (cont.)Experimental results (cont.) Knee images:– Training: 30 examples, 42 landmarksExperimental results (cont.) Search results after initial, 2 iterations, and convergence:Constrained AAMs Model results rely on starting approximation Want a method to improve influence from starting approximation Incorporate priors/user input on unseen image– MAP formulationConstrained AAMs Assume:– Gray-scale errors are uniform gaussian with variance– Model parameters are gaussian with diagonal covariance– Prior estimates of some of the positions in the image along with covariancesConstrained AAMs (cont.) We get update equation:where:Constrained AAMs Comparison of constrained and unconstrained AAM searchConclusions Combined Appearance Models
View Full Document