Make Your Conference Secure: A Robust Method to Replace Private Backgrounds

The visual conference has become one of the important parts of our daily life as the result of the worldwide pandemic. People will attend the increasing number of visual conferences with their colleagues, co-workers, or even strangers, so they are more and more concerned about how to protect their privacy when they have these meetings at home or in other personal places. To dispel people’s misgivings, the most efficient way is to hide the environmental information around attendees, or in other words, is to replace the background containing the private information with some online pictures. So based on this motivation, we propose a robust method to help people protect their private information by replacing the background with some non-privacy pictures. The proposed method, Portrait Segmentation Network(PSN), can efficiently use internet images to substitute the private background. In addition, the lack of a suitable dataset for segmentation in the visual conference scenario is another challenge. Therefore, we utilize a training method to train our proposed model by using images instead of videos in training. Meanwhile, we use the adequate experiments on EG1800 [1] and a new dataset formed by Supervisely [2] to prove the robustness of our proposed method.