Aug 22 (Thu) @ 2:00pm: "Enhanced Inter-Frame Prediction and Secondary Transform for Efficient Video Compression," Kruthika Koratti Sivakumar, ECE PhD Defense

Abstract

State-of-the-art video codecs consist of intra/inter frame prediction, transform and quantization followed by lossless entropy coding. This work focuses on the predictive and transform coding parts of a codec. The first research area explores inter-frame prediction for stereoscopic and omnidirectional videos. Current methods project spherical videos onto flat surfaces and utilize 2D video compression standards, which are suboptimal for these formats. To address this, a geodesic motion model is proposed for inter-view prediction in stereoscopic 180-degree videos, effectively capturing disparity between views. Additionally, we employ transform domain temporal prediction to improve temporal prediction for monoscopic spherical videos, overcoming limitations of pixel-domain methods. The instability issues in standard iterative closed-loop design techniques are overcome using the Asymptotic Closed-Loop (ACL) paradigm.

The second research area aims to enhance transform coding for standard 2D videos. We achieve better decorrelation and energy compaction of inter-prediction residuals by implementing non-separable secondary transforms. To address the limitations of the default coefficient scanning order in the final block of transform coefficients, we reorder these coefficients by decreasing variance, facilitating more efficient entropy coding. Additionally, a novel secondary transform coefficient selection method is developed. This algorithm selects coefficients based on the energies and mutual correlations of the primary transform coefficients. These techniques, along with the use of the ACL design paradigm, enable significant improvements in coding efficiency compared to existing standards.

Bio

Kruthika Sivakumar is a PhD candidate in the Signal Compression Lab within the Department of Electrical and Computer Engineering, under the supervision of Prof. Kenneth Rose. She received her B.E. in Electronics and Communication Engineering from Anna University, Chennai, India, in 2015, and her M.S. in Electrical and Computer Engineering from the University of California, Santa Barbara, in 2017. Her research interests include video compression, transform coding, linear predictive coding, and coding for spherical and stereoscopic videos.

Hosted by: Professor Kenneth Rose

Submitted by: Kruthika Sivakumar <kruthika@ucsb.edu>