Combining subcomponent models for object image modeling
Aspects of the present invention include systems and methods for forming generative models, for utilizing those models, or both. In embodiments, an object model fitting system can be developed comprising a 3D active appearance model (AAM) model. The 3D AAM comprises an appearance model comprising a set of subcomponent appearance models that is constrained by a 3D shape model. In embodiments, the 3D AAM may be generated using a balanced set of training images. The object model fitting system may further comprise one or more manifold constraints, one or more weighting factors, or both. Applications of the present invention include, but are not limited to, modeling and/or fitting face images, although the teachings of the present invention can be applied to modeling/fitting other objects.
1. A computer-readable medium carrying one or more sequences of instructions, wherein execution of the one or more sequences of instructions by one or more processors causes the one or more processors to form an object model fitting system based on a set of training images by performing the steps comprising:
forming a three-dimensional (3D) object shape model based on the set of training images;
forming an object appearance model using the set of training images, wherein the object appearance model comprises a linear combination of a set of object subcomponent appearance models; and
forming the object model fitting system comprising an object model comprising the object appearance model and the 3D object shape model.
2. The computer-readable medium of claim 1 wherein the object model is a face model and the training images are face images.
3. The computer-readable medium of claim 2 wherein the object model is a 3D active appearance model (AAM).
4. The computer-readable medium of claim 2 wherein the step of forming the 3D object shape model based on the set of training images comprises:
generating a set of two-dimensional (2D) shapes corresponding to a set of labeled feature points in the set of training images;
generating a set of 3D shapes by applying a factorization method to the aligned 2D shapes; and
forming the 3D object shape model by obtaining a set of orthogonal bases using the 3D shapes.
5. The computer-readable medium of claim 4 wherein the factorization method is non-rigid structure from motion.
6. The computer-readable medium of claim 5 wherein at least one symmetry constraint is applied when generating the set of 3D shapes.
7. The computer-readable medium of claim 6 wherein the at least one symmetry constraint is associated with three-dimensional depth.
8. The computer-readable medium of claim 4 wherein the set of orthogonal bases are generated using Principal Component Analysis (PCA).
9. The computer-readable medium of claim 2 wherein a subcomponent is a face subcomponent corresponding to a region of a face.
10. The computer-readable medium of claim 2 wherein the set of object subcomponent appearance models comprises a global face object appearance model.
11. A computer-readable medium carrying one or more sequences of instructions, wherein execution of the one or more sequences of instructions by one or more processors causes the one or more processors to fit an input image using an object model fitting system comprising a three-dimensional (3D) active appearance model (AAM) comprising a 3D object shape model and an object appearance model comprising a set of object subcomponent appearance models, by performing the steps comprising:
[a] generating a projected object image estimate by projecting a current 3D object shape and pose estimate obtained from the 3D object shape model onto the input image;
[b] generating an object appearance model estimate obtained from the object appearance model and appearance parameters;
[c] calculating a set of fitting errors representing a difference between the modeled object appearance estimate and the object appearance model estimate; and
[d] responsive to a stop condition not being satisfied, using at least some of the set of fitting errors to form an updated current 3D object shape and pose estimate by updating 3D shape parameter values and pose parameter values and to updated the appearance parameters, and iterating by returning to step [a].
12. The computer-readable medium of claim 11 wherein the input image is an image comprising a face and the 3D AAM models shape and appearance deformations.
13. The computer-readable medium of claim 12 wherein the set of fitting errors comprise registration errors, and wherein fitting the input image to the 3D AAM is performed by an inverse compositional method comprising finding an optimal 3D shape parameter update.
14. The computer-readable medium of claim 13 wherein the optimal 3D shape parameter update is determined by the Levenberg-Marquardt steepest descent algorithm.
15. A system for forming a reconstructed object image from an input image comprising pixels, the system comprising:
an object model fitter, coupled to receive the input image, that forms the reconstructed object image by fitting the input image using an object model fitting system comprising a three-dimensional (3D) image active appearance model (AAM), the 3D AAM comprising an object appearance model comprising a set of object subcomponent appearance models constrained by a 3D object shape model.
16. The system of claim 15 , wherein the input image comprises a face image and the 3D AAM models shape and appearance deformations across faces.
17. The system of claim 16 , wherein the object model fitter comprises:
a reconstructed object estimator, coupled to receive the input image, the 3D AAM, 3D shape parameter values, pose parameter values, and appearance parameter values, that generates a first image using the 3D object shape model, the input image, the shape parameter values, and the pose parameter values, and that generates a second image obtained using the object appearance model and appearance parameter values; and
a reconstructed object updater, coupled to receive the first and second images, that calculates a set of fitting errors representing a difference between the first and second images, and calculates updates to at least some of the 3D shape parameter values, the pose parameter values, and the appearance parameter values using at least some of the fitting errors.
18. The system of claim 17 further comprising an object model builder, coupled to receive a set of training images, the object model builder forming the 3D AAM based on the set of training images.
19. The system of claim 18 wherein the object model builder comprises:
a 3D shape reconstructor and modeler, coupled to receive a set of shapes obtained from a set of labeled feature points in the set of training images, the 3D shape reconstructor and modeler forming the 3D object shape model based on the set of shapes; and
a subcomponent appearance modeler, coupled to receive the set of training images, the subcomponent appearance modeler forming the set of object subcomponent appearance models using the set of training images.
20. The system of claim 19 wherein the set of shapes are two-dimensional (2D) shapes, and forming the 3D object shape model based on the set of shapes comprises:
aligning the 2D shapes;
generating 3D shapes by applying a factorization method to the aligned 2D shapes; and
forming the 3D object shape model by obtaining a set of orthogonal bases using the 3D shapes.