Tag Archive for: university klagenfurt

,

IEEE ICIP’25: Machine Learning-Based Decoding Energy Modeling for VVC Streaming

Machine Learning-Based Decoding Energy Modeling for VVC Streaming

2025 IEEE International Conference on Image Processing (ICIP)

14-17 September, Anchorage, Alaska, USA

https://2025.ieeeicip.org/

Reza Farahani (AAU Klagenfurt, Austria), Vignesh V Menon (Fraunhofer HHI, Germany), and Christian Timmerer (AAU Klagenfurt, Austria)

Abstract: Efficient video streaming requires jointly optimizing encoding parameters (bitrate, resolution, compression efficiency) and decoding constraints (computational load, energy consumption) to balance quality and power efficiency, particularly for resource-constrained devices. However, hardware heterogeneity, including differences in CPU/GPU architectures, thermal management, and dynamic power scaling, makes absolute energy models unreliable, particularly for predicting decoding consumption. This paper introduces the Relative Decoding Energy Index (RDEI), a metric that normalizes decoding energy consumption against a baseline encoding configuration, eliminating device-specific dependencies to enable cross-platform comparability and guide energy-efficient streaming adaptations. We use a dataset of 1000 video sequences to extract complexity features capturing spatial and temporal variations, employ Versatile Video Coding (VVC) open-source toolchain using VVenC/VVdeC with various resolutions, framerate, encoding preset and quantization parameter (QP) sets, and model RDEI using Random Forest (RF), XGBoost, Linear Regression (LR), and Shallow Neural Networks (NN) for decoding energy prediction. Experimental results demonstrate that RDEI-based predictions provide accurate decoding energy estimates across different hardware, ensuring cross-device comparability in VVC streaming.

Keywords: Video Streaming; Energy Prediction; Versatile Video Coding (VVC); Video Complexity Analysis.

 

 

 

 

 

,

Paper accepted: ICME 2025: Neural Representations for Scalable Video Coding

Neural Representations for Scalable Video Coding

IEEE International Conference on Multimedia & Expo (ICME) 2025

 

Authors: Yiying Wei (AAU, Austria), Hadi Amirpour (AAU, Austria),  and Christian Timmerer (AAU, Austria)

 

Abstract: Scalable video coding encodes a video stream into multiple layers so that it can be decoded at different levels of quality/resolution, depending on the device’s capabilities or the available network bandwidth. Recent advances in implicit neural representation (INR)-based video codecs have shown competitive compression performance to both traditional and other learning-based methods. In INR approaches, a neural network is trained to overfit a video sequence, and its parameters are compressed to create a compact representation of the video content. While they achieve promising results, existing INR-based codecs require training separate networks for each resolution/quality of a video, making them challenging for scalable compression. In this paper, we propose Neural representations for Scalable Video Coding (NSVC) that encodes multi-resolution/-quality videos into a single neural network comprising multiple layers. The base layer (BL) of the neural network encodes video streams with the lowest resolution/quality. Enhancement layers (ELs) encode additional information that can be used to reconstruct a higher resolution/quality video during decoding using the BL as a starting point. This multi-layered structure allows the scalable bitstream to be truncated to adapt to the client’s bandwidth conditions or computational decoding requirements. Experimental results show that NSVC outperforms AVC’s Scalable Video Coding (SVC) extension and surpasses HEVC’s scalable extension (SHVC) in terms of VMAF. Additionally, NSVC achieves comparable decoding speeds at high resolutions/qualities.

 

ICME 2025: Neural Representations for Scalable Video Coding | ATHENA Christian Doppler (CD) Laboratory