Core Concepts
GRACE achieves enhanced loss resilience in real-time video communication through joint training of neural encoder and decoder.
Abstract
Abstract:
Retransmitting lost packets in real-time video communication is not feasible due to latency requirements.
Two primary strategies for countering packet losses are encoder-based FEC and decoder-based error concealment.
GRACE introduces a loss-resilient real-time video system using a new neural video codec.
Introduction:
Real-time video communication is crucial for various applications, necessitating protection against packet losses.
Loss-Resilient Techniques:
Encoder-side FEC adds redundancy before transmission, while decoder-side error concealment reconstructs lost data.
GRACE's Approach:
Jointly optimizing the encoder and decoder under simulated packet losses enhances loss resilience.
Training Process:
Simulating packet losses during training improves the NVC's ability to handle diverse loss rates effectively.
Streaming Protocol:
Optimistic encoding and dynamic state resynchronization prevent out-of-sync states without blocking encoding or decoding processes.
Fast Coding and Bitrate Control:
GRACE-Lite optimizes NVC for efficient execution on CPUs and mobile devices, achieving accurate bitrate control.
Stats
GRACE achieves a 38% higher mean opinion score (MOS) than other baselines in a user study with 240 participants.
GRACE reduces undecodable frames by 95% compared to FEC.
GRACE accelerates encoding and decoding by 4× without impacting loss resilience.