Video Retargeting via Sound Localization and Preserving Maps

The patent applies local quality by creating a preserving map that assigns different importance weights to different regions of the video frame. Important regions (active speaker, body language) are assigned higher weights to be preserved, while less important regions are assigned lower weights and can be distorted or cropped. This allows selective preservation of critical information while maintaining overall field of view.

The patent segments the video frame into multiple independent regions of interest rather than using a single rectangular crop. Each detected person or important object becomes a separate region with its own bounding box and importance weight. This allows the system to handle multiple persons of interest simultaneously while keeping the detection algorithm relatively simple.

The patent replaces static linear scaling with dynamic nonlinear scaling that adapts to content importance. The preserving map guides a warp function that dynamically adjusts the scaling factor for each region based on its importance weight. This allows the system to fill the entire display area without black borders while preserving important regions, making the field of view adaptable rather than fixed.

The patent substitutes the mechanical conversion process (raster to vector) with a direct raster image processing approach. Instead of converting raster video to vector format and then applying content-aware resizing, the method works directly on raster images using a preserving map to guide nonlinear scaling. This avoids the quality degradation associated with vector conversion while still achieving content-aware resizing on natural videos.