Shape Matching and Registration by Data-driven EM

Area of Science:

• Computer Vision
• Geometric Algorithms

Background:

• Shape matching, registration, and detection are crucial in computer vision.
• Existing generative models (e.g., EM algorithm) face initialization and E-step computation challenges.
• Sparse-point and continuous-contour representations are common for shape data.

Purpose of the Study:

• To develop an efficient and robust algorithm for shape matching, registration, and detection.
• To overcome limitations of traditional generative models by incorporating shape features.
• To achieve rapid convergence and high accuracy in geometric transformations.

Main Methods:

• Formulated shape matching as probabilistic inference using a generative model and EM algorithm.
• Developed a discriminative model leveraging shape features to address generative model limitations.
• Created a hybrid algorithm combining both generative and discriminative approaches.
• Utilized sparse-point or continuous-contour shape representations.

Main Results:

• The hybrid algorithm demonstrates high efficiency and robustness in shape matching and registration.
• Shape features significantly improve correspondence solving, requiring typically only four iterations.
• Algorithm convergence time is consistently under one second.
• Successful validation on standard datasets like MPEG7 and diverse segmentation tasks.

Conclusions:

• The hybrid generative-discriminative algorithm offers a significant advancement in shape analysis.
• The method provides a fast and accurate solution for geometric transformation tasks.
• This approach is effective for various applications including shape matching, registration, and segmentation.