Region-of-Interest-Based Video Coding for Machines

The paper is within the scope of Video Coding for Machines (VCM), where video compression not only caters to human viewers but also acts as input to systems engaged in diverse machine vision tasks like object detection and tracking. A novel compression technique is proposed, which exemplifies the proficient utilization of Rols-based video coding. It includes significant modifications to the state-of-the-art technique employing image retargeting and also addresses the related problem of resolution change. In the novel approach presented in this work, due to the proposed application of image padding, video frames encoded by an Inner Encoder (e.g. a VVC encoder) remain of constant size. Experimental evaluation demonstrates significant average bitrate reduction, up to 57% with respect to the current version of VCM, while maintaining consistent quality across various machine vision tasks and encoding scenarios.