Multi-View AAM Fitting and Construction.

Area of Science:

• Computer Vision
• Computer Graphics
• Machine Learning

Background:

• Active Appearance Models (AAMs) are established generative models for deformable objects like human faces.
• Existing AAMs formulations are primarily 2D, with recent extensions to 3D shape models.
• Current 3D AAM fitting and construction methods predominantly rely on single-view algorithms, leaving multi-view approaches underexplored.

Purpose of the Study:

• To develop and present novel multi-view algorithms for both fitting and constructing 3D Active Appearance Models (AAMs).
• To investigate and improve the accuracy and robustness of 3D AAMs through multi-view data integration.
• To address limitations in existing 3D AAM construction, specifically the Bas-Relief ambiguity.

Main Methods:

• Developed a multi-view algorithm for fitting a single 3D AAM to multiple simultaneously captured images, including automatic camera calibration (scaled orthographic and weak/full perspective models).
• Utilized the human face as a non-rigid calibration grid for camera model calibration.
• Proposed a robust non-rigid motion-stereo algorithm for calibrated multi-view 3D AAM construction.

Main Results:

• Demonstrated that the proposed camera calibration algorithm achieves performance comparable to standard calibration grid methods.
• Showcased that camera calibration significantly enhances the performance of AAM fitting.
• Proved that calibrated multi-view motion-stereo effectively eliminates the Bas-Relief ambiguity, leading to higher fidelity 3D face models.

Conclusions:

• Multi-view algorithms offer significant advantages for fitting and constructing 3D Active Appearance Models.
• Integrated camera calibration within the multi-view framework improves the accuracy of 3D AAM fitting.
• Calibrated multi-view motion-stereo is a robust solution for high-fidelity 3D face model construction, overcoming inherent ambiguities.