Dejavu: Enhancing videoconferencing with prior knowledge

Abstract

Videoconferencing over the Internet routinely suffers from poor quality as videoconferencing systems, in order to guarantee interactive delays which is critical to user experience, are commonly designed to stream at conservative qualities in the face of variable bandwidths. In this paper, we present Dejavu, a system that enables existing videoconferencing systems to alleviate this problem. The key insight that powers Dejavu is that recurring videoconferencing sessions, e.g., in the same conference room or by the same person, have a lot of visual similarities that can be encoded based on the sender's historical videoconferencing sessions, and shared with the receiver in advance. Accordingly, Dejavu first learns an offline mapping between low-quality and high-quality versions of frames in the sender's past videoconferencing sessions, and then applies this mapping in real time at the receiver to convert the low-quality frames into high-quality frames. As a result, a videoconferencing system equipped with Dejavu can continue to stream at conservative qualities to guarantee interactive delays like today, but can now additionally enhance the video quality at the receiver. Our evaluation shows that Dejavu can provide a 1.3 dB increase in PSNR for the same bandwidth consumption, or equivalently save up to 30% in bandwidth to deliver the same PSNR.