Scale Space for Camera Invariant Features.

Area of Science:

• Computer Vision
• Differential Geometry
• Image Processing

Background:

• Scale space representation is crucial for image analysis tasks like feature detection (e.g., SIFT).
• Existing methods often require camera-specific calibration or are limited to certain projection models.
• Central projection systems (catadioptric, fisheye, perspective) exhibit unique geometric properties.

Purpose of the Study:

• To develop a unified framework for computing scale space applicable to any central projection system.
• To enable automatic computation of the Riemannian metric and Laplace-Beltrami operator for diverse camera models.
• To demonstrate the effectiveness and generalizability of the proposed scale space computation method.

Main Methods:

• A unified mathematical model for central projection systems.
• Automatic computation of the Riemannian metric based on a single system parameter.
• Application of the partial differential equations framework on manifolds.
• Computation of the Laplace-Beltrami (LB) operator for scale space generation.

Main Results:

• Successfully computed the scale space for various central projection systems (catadioptric, fisheye, perspective).
• Validated the approach using both synthetic and real-world image data.
• Demonstrated competitive performance compared to existing state-of-the-art methods across different camera types.

Conclusions:

• The proposed method offers a generalized solution for scale space computation in computer vision.
• This unified approach simplifies feature extraction and scale selection for a wide range of imaging systems.
• The method shows significant potential for improving performance in various image analysis applications.