May 27 (Thu): "Towards Optimal Compression Paradigms," Bharath Vishwanath, ECE PhD Defense
Video forms an important component of the data consumption and constitutes a major part of the internet traffic. Evidently, there is a pressing need for better compression algorithms. A recent challenge to the research community is the compression of spherical videos that forms the backbone of many virtual reality related applications. In this talk we present novel compression algorithms for spherical and 2D videos.
First, we consider compression of an important class of spherical videos with dominant camera motion. With camera motion, we observe that the surrounding static objects are perceived to move along geodesics on the sphere paving way for a novel geodesic translation motion model on the sphere. We then present our analysis that relates the rate of translation of pixels to the depth of the pixels and elevation of pixels on the sphere. This enables us to accurately capture the motion of each pixel on the sphere.
In the second part of the talk, we present our work on prediction filter design for adaptive signal compression with applications to video coding. The design poses several challenges including design instability due to closed-loop nature of the coders and the severe non-convexity of the cost function. We present a deterministic annealing based asymptotic closed-loop design paradigm that effectively addresses all the design challenges.
The final part of the talk considers the design of transforms for inter-prediction residual in video coding. The design instability due to closed-loop nature of the codec is a major stumbling block that severely hinders transform optimization. We present an asymptotic closed-loop design strategy that gives a stable platform to optimize transforms. The proposed methods are shown to significantly outperform the existing codecs and other state-of-the art approaches.
Bharath Vishwanath received B.E degree in Electronics and Communications Engineering from the National Institute of Technology Karnataka, India, in 2014 and the M.Sc. in ECE from the University of California, Santa Barbara (UCSB), in 2016. He is currently a doctoral candidate in Signal Compression Lab, UCSB. His research interests include video coding, non convex optimization and information theory.
Hosted by: Professor Kenneth Rose
Submitted by: firstname.lastname@example.org