Active Appearance Models: Theory, Extensions and Cases |
Mikkel Bille Stegmann
|
| Abstract | This thesis presents a general approach towards image segmentation using the learning-based deformable model Active Appearance Model (AAM) proposed by Cootes et al. The primary advantage of AAMs is that a priori knowledge is learned through observation of both shape and texture variation in a training set. From this, a compact object class description is derived, which can be used to rapidly search images for new object instances.
A thorough treatment and discussion of the theory behind AAMs is given, followed by several extensions to the basic AAM, which constitutes the major contribution of this thesis. Extensions include automatic initialization and unification of finite element models and AAMs. All of these have been implemented in a structured and fast C++ framework; the AAM-API.
Finally, case studies based on radiographs of metacarpals, cardiovascular magnetic resonance images and perspective images of pork carcass are presented. Herein the performance of the basic AAM and the developed extensions are assessment using leave-one-out evaluation.
It is concluded that AAMs -- as a data-driven and fully automated method -- successfully can perform object segmentation in challenging and very different image modalities with very high accuracy. In two of three cases subpixel accuracy were obtained w.r.t. object segmentation. |
| Type | Master's thesis [Academic thesis] |
| Year | 2000 Month August pp. 262 Ed. 2 |
| Publisher | Informatics and Mathematical Modelling, Technical University of Denmark, DTU |
| Address | Richard Petersens Plads, Building 321, DK-2800 Kgs. Lyngby |
| IMM no. | IMM-EKS-2000-25 |
| Electronic version(s) | [pdf] |
| Publication link | http://www.imm.dtu.dk/~aam/main/ |
| BibTeX data | [bibtex] |
| IMM Group(s) | Image Analysis & Computer Graphics |