(Privacy Preserving) Visual Localization | CESCG 2023 Keynote

Visual localization is the problem of estimating the precise position and orientation from which a given image was taken. Solving the visual localization step is a fundamental component of technologies such as Augmented and Virtual Reality systems and autonomous robots such as self-driving cars. After an introduction to the visual localization problem, including an overview over different solution strategies for the localization problem, this talk focuses on the sub-field of privacy-preserving localization. Given the rising need for and availability of cloud-based localization services (e.g., offered by Google, Microsoft, and Niantic), privacy-preserving localization approaches aim to ensure that no private details can be recovered from user data shared with such services. This data is typically in the form of images or user-generated 3D maps. We will give a brief overview over existing privacy-preserving approaches and will then take a critical look at whether these approaches are as privacy-preserving as advertised.

The keynote was given by Torsten Sattler from Czech Technical University in Prague.

00:00 Intro
01:15 What is visual localization
03:50 Camera pose regression
15:09 Structure-based localization
18:50 Overview of other methods
21:24 Privacy aspects
26:18 Hardening input data using 3D lines
34:15 Reverse engineering of output poses
38:56 Conclusions
40:20 Q&A: What damage might occur when my data is compromised?
42:39 Q&A: Is localization without a cloud server possible?
45:35 Q&A: Are robustness, accuracy and privacy antitheses?
47:25 Q&A: What about privacy protecting scrambling of 3D Scenes?
49:04 Q&A: Is the output data exploitation relevant at all?
50:38 Q&A: Can offline visual localization work in indoor environments filled with visitors?
53:14 Q&A: Is the idea of PGP applicable to online localization?
55:19 Q&A: Would it be possible to learn descriptors that would introduce ambiguous values to preserve privacy?