Divya Guruswamy

Semester Work
Supervisors: Dr. Danda Pani Paudel and Prof. Luc Van Gool

On Privacy preserving Data Obfuscation in Two-View Geometry

To hide the sensitive information in images, key-point feature locations and descriptors are shared for various applications of the two-view geometry. However, recent state-of-the-art methods have shown that the original image can be reconstructed faithfully, form the shared features alone, thus leading to the privacy risk. This demands a privacy preserving encryption, which also allows estimating two-view relationships, so that the users are encouraged to share images for two-view applications. In this work, we propose a method to preserve the privacy between two users, each capturing one image of two-views. The propose method obfuscates the image points by sharing a projectively ambiguous linear system of equations. The coefficients of these equations are derived from the measurements of the first user, whereas, the variables represent the unknown measurements of the second user. When shared by the first user, the second user derives a constant measurement matrix, whose properties are used to derive the two-view relationships. This simple method hides the original points from the second user, irrespective of the number of obfuscated points matched, while still making point matching and validation feasible. In this thesis, we analyses several theoretical aspects of the developed method, in terms of data sharing, relative pose estimation, and the user privacy.