Slimmable Video Codec
This addresses the need for more practical and adaptable neural video compression for applications requiring efficient resource usage, though it is incremental by extending slimmable techniques from images to video.
The paper tackles the impracticality of neural video codecs due to heavy architectures and fixed rate-distortion tradeoffs by proposing a slimmable video codec that dynamically adjusts model capacity, showing it effectively controls rate, memory, computation, and latency without harming performance.
Neural video compression has emerged as a novel paradigm combining trainable multilayer neural networks and machine learning, achieving competitive rate-distortion (RD) performances, but still remaining impractical due to heavy neural architectures, with large memory and computational demands. In addition, models are usually optimized for a single RD tradeoff. Recent slimmable image codecs can dynamically adjust their model capacity to gracefully reduce the memory and computation requirements, without harming RD performance. In this paper we propose a slimmable video codec (SlimVC), by integrating a slimmable temporal entropy model in a slimmable autoencoder. Despite a significantly more complex architecture, we show that slimming remains a powerful mechanism to control rate, memory footprint, computational cost and latency, all being important requirements for practical video compression.