ECCV 2022 - Wednesday

to this lens, ” he explains. “ One of the biggest challenges was optimizing the network because our approach started with models unconcerned about privacy. We began pre-training models on computer vision tasks, and as we fine-tuned, there was a trade-off between privacy and task performance. ” The team tried two human action recognition networks: C3D and Rubiksnet . With C3D, it was difficult to achieve this trade-off, but Rubiksnet showed better results with modifications, including temporal similarity matrices to preserve the temporal information that could be lost when the input video is distorted. There are several potential applications for this work. In hospitals, where cameras perform vital computer vision tasks, this model could help preserve patients’ privacy, with the added benefit of enabling the collection of anonymized patient data that could be used for further research. It could also be used at home when a family may wish to monitor an elderly relative’s activity to know if they have fallen, for example, without intruding on their privacy. “ We had a previous paper that designed a camera to preserve privacy for human pose estimation, ” Carlos adds. “ That could also be implemented in surgical rooms to monitor the movement of patients and doctors. ” This work was conceived when the team met virtually during the height of the pandemic to discuss designing a privacy-preserving system. It’s been a true collaboration, with Henry Arguello specializing in optics and 5 Carlos Hinojosa