Authors: Axel Tobieson Rova & Jonatan Kruse

Supervisors: Ivar Persson, Robin Gustavsson (Spiideo), Håkan Ardö (Spiideo)

Examiner: Mikael Nilsson, CVML

Abstract:

Modern high-resolution cameras generate enormous amounts of video data, creating significant challenges for storage and transmission in bandwidth-constrained environments. This thesis investigates region-based video compression techniques that intelligently allocate bits to optimize perceived quality in sports video content, where visual importance varies  significantly across the frame.

Two primary approaches were developed and evaluated: spatial pixel transformations that redistribute pixel density toward regions of interest before encoding, and quantization-based methods that apply varying compression levels to different regions during the encoding process. The research utilized Spiideo’s sports video pipeline, incorporating field boundary data and action area information to guide region prioritization. The quantization-based approach demonstrated substantial improvements in perceptual quality for high-priority regions. Spatial transformations also achieved quality gains, though generally underperformed compared to quantization methods. Both techniques successfully preserved quality in critical areas while allowing controlled degradation in less important regions such as audience areas and out-of-field content.

Comprehensive evaluation across multiple viewports using VMAF, SSIM, and PSNR metrics validated the effectiveness of region-aware compression strategies. The research demonstrates that intelligent bit allocation can effectively bridge the gap between modern camera capabilities and network transmission constraints, preserving high-quality capture where it provides the greatest perceptual value to viewers in sports video applications.