1. Introduction

Wearable cameras are used as a tool to understand fine-grained human activities in the wild because of their ability to provide visual information that can be interpreted by humans [15, 45, 55] or machines [6, 43, 48]. Particularly in the ubiquitous computing (UbiComp) community, wearable cameras are increasingly being used to obtain visually confirmed annotations of wearers’ activities in real-world settings, which is necessary to both understand human.

Behavior at a fine-grained level, and build and validate non-visual wearable devices and their corresponding supervised machine learning algorithms to automate the detection of human activity [4, 8, 9, 61, 80]. However, the stream of images obtained from these wearable cameras embeds more details than needed

RQ1: How do different activity-oriented partial obfuscation filters affect the visual confirmation utility of identifying hand-related activities that involve hand-to-head gestures by a human viewer?

In particular, we want to compare the accuracy of human labels obtained from viewing non-obfuscated videos with the accuracy of the labels derived from viewing the obfuscated videos with different filters. Hand-to-head gestures can be confounding to each other if fine-grained and some contextual information is lost. Therefore, this comparison can help us to determine if the visual confirmation utility is preserved, or not, after applying activity-oriented partial obfuscation to it with different filters. It will also help us to understand the limitations of activity-oriented partial obfuscation and the filters applied.

RQ1: How do different activity-oriented partial obfuscation filters affect the visual confirmation utility of identifying hand-related activities that involve hand-to-head gestures by a human viewer?

In particular, we want to compare the accuracy of human labels obtained from viewing non-obfuscated videos with the accuracy of the labels derived from viewing the obfuscated videos with different filters. Hand-to-head gestures can be confounding to each other if fine-grained and some contextual information is lost. Therefore, this comparison can help us to determine if the visual confirmation utility is preserved, or not, after applying activity-oriented partial obfuscation to it with different filters. It will also help us to understand the limitations of activity-oriented partial obfuscation and the filters applied.