TASK-AWARE UNIFIED DOWNSCALING FOR BANDWIDTH-EFFICIENT TELEMETRY VIDEO TRANSMISSION

Abstract

In this paper, we present a task-aware unified downscaling framework applicable to telemetry video systems, which determines the optimal downscaling factor based on a Rate-Loss optimization criterion for efficient video transmission. Although prior studies have considered the trade-off among compression rate, task accuracy, and latency, most methods adopt identical scaling factors for both width and height or rely on deep learning–based modules that require per-image optimization, leading to high computational cost. The proposed method introduces a Task-Aware Unified Downscaling Factor (TUDF) that determines an optimal scale factor minimizing bitrate while maintaining detection accuracy. By eliminating the need for a deep learning–based downscaling module, the approach substantially reduces computational overhead and inference latency, thereby offering a practical solution for real-time telemetry video transmission. Experimental results demonstrate that the proposed approach can significantly reduce bitrate while preserving the performance of downstream computer vision tasks.