Video compression technique using facial landmarks on mobile devices

Abstract

Wide proliferation of smart mobile phones has manifolded the bandwidth demand, as video streaming applications have significantly gained popularity. At the same time, technical challenges, such as requirements of resources, as well as practical challenges such as limited availability of the mobile bandwidth spectrum, have acted as inhibitors of transmission of unlimited video over the wire in an ubiquitous manner. In this thesis, we propose a first-of-its-kind methodology for compressing videos that stream human faces. Our technique is amenable for streaming transmission of live videos. Our framework relies upon detecting facial landmarks on-the-fly, and compressing the video by storing a sequence of distinct frames extracted from the video, such that the facial landmarks of a pair of successively stored frames are significantly different. The compression technique uses a dynamic thresholding technique to detect significance of difference, and stores meta-information for reconstructing the missing frames. We measure the goodness of our technique by evaluating the time taken to compress, the entropy of successively stored images, and a comparison with several static thresholds of significance. We validate our work with a user study, observing user satisfaction at different compression ratios. Our work will also be useful in applications that require live streaming of facial videos.