Learning-Based Approach to Estimation of Morphable Model Parameters
Vinay P. Kumar, Tomaso Poggio · DSpace@MIT (Massachusetts Institute of Technology) · 2000
The Problem: Developing a method for estimating the parameters of a linear morphable model (LMM) directly from the images of the object class, based on a learning from examples approach. Motivation: Amongst the many model-based approaches to modeling object classes, the Linear Morphable Model is an important one (Vetter and Poggio [6], Jones and Poggio [2]). It has been been used successfully to model faces, cars and digits. In these applications, the task of matching a novel image to the LMM is achieved through a computationally intensive analysis by synthesis approach. In Jones and Poggio [2], the matching parameters are computed by minimizing the squared error between the novel image and the model image using a stochastic gradient descent algorithm. This technique may take several minutes for matchingevenasingle image. A technique that could compute the matching parameters with considerably less computations and using only view-based representations would make these models useful in real-time applications. The motivation for this work comes from the use of a learning-based approach in real-time analysis of mouths (Kumar and Poggio [4]), in which it was shown that a regression function can be learnt from a Haar wavelet based input representation of mouths to hand labeled parameters denoting openness and smile. Therefore, it points to the possibility that learning may be a way for directly estimating the matching parameters of an LMM from the image. Previous Work: Previously, morphable models of mouths have been constructed for the purpose of synthesis of