Error-Resilience Profiling of Inter-Frame Prediction at VVC Encoders for Approximate Storage

This paper presents an error tolerance (aka. resilience) profiling, focusing on the inter-frame prediction at Versatile Video Coding (VVC) encoders, for approximate storage exploitation. The goal of this work is to investigate the feasibility of using approximate storage in three specific memory regions during the inter-frame prediction step using VVC: Original, Filtered and Prediction Samples Buffers. The study evaluated the resulting losses in terms of coding efficiency by injecting random errors in four video sequences. The results indicate that increased error rate results in higher Bjontegaard Delta-Rate (BD-Rate) values. As insights from the performed analysis, it was observed a great potential for exploring approximate storage, since the increase in BD-Rate did not exceed 0.9% and the technique proved to be more efficient on average was the Prediction Samples Buffer (PredSB) memory, followed by the Original Samples Buffer (OrigSB) and Filtered Samples Buffer (FiltSB). Furthermore, it was observed that different video characteristics lead to different levels of error resilience.