Artificial Intelligence based Video Codec (AIVC) for CLIC 2022

TL;DR

This paper introduces AIVC, a fully-learned video codec utilizing conditional autoencoders, which optimizes rate allocation and frame structure to achieve significant compression improvements in the CLIC 2022 competition.

Contribution

The paper presents a novel learned video codec that dynamically optimizes coding configurations, leading to enhanced compression efficiency over traditional methods.

Findings

Achieved 26% rate reduction compared to non-optimized configurations.

Implemented a flexible autoencoder-based model for adaptive rate allocation.

Demonstrated competitive performance in the CLIC 2022 video track.

Abstract

This paper presents the AIVC submission to the CLIC 2022 video track. AIVC is a fully-learned video codec based on conditional autoencoders. The flexibility of the AIVC models is leveraged to implement rate allocation and frame structure competition to select the optimal coding configuration per-sequence. This competition yields compelling compression performance, offering a rate reduction of -26 % compared with the absence of competition.