Enabling privacy-preserving first-person cameras using low-power sensors

Wu, Muchen; Pathak, Parth H.; Mohapatra, Prasant

doi:10.1109/sahcn.2015.7338345

Cited by 4 publications

(2 citation statements)

References 16 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…The authors reported an accuracy of 70% during testing on 300 images. Muchen et al [25], proposed a privacy-preserving approach for the head-mounted wearable sensor using multi-sensor information from a mobile phone and smartwatch. Here the authors used the multi-sensor information to identify privacy concern scenarios and trigger the wearable camera.…”

Section: Privacy Content Removal Methodsmentioning

confidence: 99%

Selective Content Removal for Egocentric Wearable Camera in Nutritional Studies

Hassan

Sazonov

2020

IEEE Access

View full text Add to dashboard Cite

Research and innovation on wearable sensor technology has grown exponentially over this decade. The privacy-related concerns have grown in tandem. Automatic Ingestion Monitor v2 (AIM-2) is an egocentric camera and sensor that aids monitoring of individual diet and eating behavior by capturing still images throughout the day and using sensor data to detect eating. The images may be used to recognize foods being eaten, eating environment, and other behaviors and daily activities. At the same time, captured images may carry privacy concerning content such as (1) people in social eating and/or bystanders (i.e., bystander privacy); (2) sensitive documents that may appear on a computer screen in the view of AIM-2 (i.e., context privacy). Managing privacy concerns by discarding images potentially containing private content is not practical as many activities are performed socially or with the use of computing devices. Therefore, we propose an automatic, image redaction-based privacy protection by selective content removal based on semantic segmentation using a deep learning neural network. We prioritize removal of humans and display screens. A free-living dataset containing 1 day of data from 15 participants was used to develop and validate the proposed method. We used 331,614 augmented data from 9 participates and publicly available data for training. The proposed method was validated on data of 6 participants and reported bystander privacy removal with precision of 0.87 and recall of 0.94 and reported context privacy removal by precision and recall of 0.97 and 0.98. The results of the study showed that selective content removal using deep learning neural network is a much more desirable approach to address privacy concerns for an egocentric wearable camera.

show abstract

Section: Privacy Content Removal Methodsmentioning

confidence: 99%

Selective Content Removal for Egocentric Wearable Camera in Nutritional Studies

Hassan

Sazonov

2020

IEEE Access

View full text Add to dashboard Cite

show abstract

“…When the target domain changes by the user's movement, adapting to multiple target domains is desired. When we cannot access the target data due to privacy issues [47], [53], [60], generalizing to an unseen domain is also necessary. We also experimented on a multitarget domain adaptation setting [2], [11], which aims to simultaneously adapt to multiple target domains, and a multiauxiliary domain generalization setting [62], which aims to generalize to an unseen test domain utilizing knowledge from auxiliary domains.…”

Section: Introductionmentioning

confidence: 99%

Foreground-Aware Stylization and Consensus Pseudo-Labeling for Domain Adaptation of First-Person Hand Segmentation

et al. 2021

View full text Add to dashboard Cite

Hand segmentation is a crucial task in first-person vision. Since first-person images exhibit strong bias in appearance among different environments, adapting a pre-trained segmentation model to a new domain is required in hand segmentation. Here, we focus on appearance gaps for hand regions and backgrounds separately. We propose (i) foreground-aware image stylization and (ii) consensus pseudolabeling for domain adaptation of hand segmentation. We stylize source images independently for the foreground and background using target images as style. To resolve the domain shift that the stylization has not addressed, we apply careful pseudo-labeling by taking a consensus between the models trained on the source and stylized source images. We validated our method on domain adaptation of hand segmentation from real and simulation images. Our method achieved state-of-the-art performance in both settings. We also demonstrated promising results in challenging multi-target domain adaptation and domain generalization settings.

show abstract