SCALED : Surrogate-gradient for Codec-Aware Learning of Downsampling in ABR Streaming

The rapid growth in video consumption has introduced significant challenges to modern streaming architectures. Over-the-Top (OTT) video delivery now predominantly relies on Adaptive Bitrate (ABR) streaming, which dynamically adjusts bitrate and resolution based on client-side constraints such as display capabilities and network bandwidth. This pipeline typically involves downsampling the original high-resolution content, encoding and transmitting it, followed by decoding and upsampling on the client side. Traditionally, these processing stages have been optimized in isolation, leading to suboptimal end-to-end rate-distortion (R-D) performance. The advent of deep learning has spurred interest in jointly optimizing the ABR pipeline using learned resampling methods. However, training such systems end-to-end remains challenging due to the non-differentiable nature of standard video codecs, which obstructs gradient-based optimization. Recent works have addressed this issue using differentiable proxy models, based either on deep neural networks or hybrid coding schemes with differentiable components such as soft quantization, to approximate the codec behavior. While differentiable proxy codecs have enabled progress in compression-aware learning, they remain approximations that may not fully capture the behavior of standard, non-differentiable codecs. To our knowledge, there is no prior evidence demonstrating the inefficiencies of using standard codecs during training. In this work, we introduce a novel framework that enables end-to-end training with real, non-differentiable codecs by leveraging data-driven surrogate gradients derived from actual compression errors. It facilitates the alignment between training objectives and deployment performance. Experimental results show a 5.19\% improvement in BD-BR (PSNR) compared to codec-agnostic training approaches, consistently across the entire rate-distortion convex hull spanning multiple downsampling ratios.