Anonymisation in egocentric video

A summary of research related to anonymisation in egocentric video.

Detailed description:
We begin by discussing the rise of egocentric video and its use-cases spanning law enforcement, extreme sports, geriatric care and life-logging. We then briefly examine the privacy implications of releasing egocentric videos and the increasing availability of egocentric research datasets.

Next, we discuss human gait, introducing early work in psychology and automatic recognition of gait using low-level cues, silhouettes and 3D accelerometers.

We then turn to more recent work, describing recognition of video photographer identity using neural networks to find correspondences between sparse optical flow features, visual matching of individuals across views, embedding architectures for deep metric learning and linking gait across different views and time. We also discuss the recognition of egocentric videographers from their hands.

Finally, we describe a perturbation-based anonymisation defence against identity attacks.

Timestamps:
0:55 - Motivation
3:11 - Gait and its significance for egocentric video anonymity
9:29 - Attack: Gait recognition
12:04 - Attack: Combining egocentric footage with third-person footage
19:50 - Attack: Hand gestures
22:01 - Defence: An anonymisation strategy for egocentric videos
26:14 - Summary

Topics: #egocentric #privacy #machinelearning #computervision

Slides (pdf): https://samuelalbanie.com/files/diges...

References: A full list of the references for the video can be found at https://samuelalbanie.com/digests/202...