Future-frame clues help neural codecs compress video
Esta página aún no tiene una traducción verificada. Se muestra la versión en inglés. Leer la versión en inglés
En palabras sencillas
Most video codecs compress each frame using frames that have already been reconstructed. That leaves out clues in nearby future frames.
Graph-Factored Sequential Coding (GFSC) is a video-compression framework that lets the encoder look ahead by a chosen number of frames and combine information from that small window.
The authors tested GFSC additions to three neural codecs on two video datasets. At similar picture quality, the enhanced codecs used less data than their causal counterparts in those tests.
The paper also proves a rate-optimality result for certain mathematical source models, within a limited range of reconstruction errors.
La investigación
Causal video codecs cannot use future-frame correlations when encoding the current frame. Could a limited look-ahead improve the amount of data needed for a given picture quality, while still fitting a sequential coding process?
GFSC gives sequential video codecs a bounded future-frame window, and proves rate-distortion optimality for specified Markov sources within limited distortion regions.
The authors test GFSC by adding its delayed-frame modules to three neural codecs and comparing each with a causal counterpart on two video datasets. Across the tested delays, the enhanced codecs consistently use less bitrate at comparable reconstruction quality.
This evidence supports the view that limited future-frame information can improve compression in these systems. A parameter-matched widened causal DCVC also performs worse than GFSC in a separate comparison, supporting delayed modeling as a contributing distinction there.
The optimality proofs apply only to specified source models and nonempty distortion regions. The measured compression gains cover three codec families, two video datasets, and delays up to three frames; training procedures were not matched in every comparison.
Análisis
Detalles
- Autores del paper original
- Likun Cai, Jun Chen, Ashish Khisti
- Enviado a arXiv
- Licencia
- CC BY 4.0