Grace: A Data Scalable Real-Time Video Communication System
SUMMARY
- Real-time video communications are essential to the daily routine of users for online meetings, cloud gaming, and interactive virtual reality. Retransmitting data packets that are not compatible can lead to stuttering, poor quality, and user frustration.
- Current technologies, such as DVC and QRAM, share similar neural network video encoders to retransmit data packets but lack an achievable rate of data-scalable video coding. These error-resilient technologies use redundancy to handle packet loss which compromises quality and compression efficiency. Moreover, these systems are disorganized as redundancy is increased even after frames have experienced packet loss.
- The faculty inventor has found a way to maintain smooth video playback without waiting for retransmitted packets. Grace, a new autoencoder-based real-time video communication system, uses neural networks in a novel packetization scheme that customizes autoencoders. Grace proves that the customization of the autoencoder training process can make video streaming loss-resilient which is an attribute that standard autoencoders lack.
- When all data is received through Grace the decoded visual quality will be on par with advanced video coding formats (H.264 and H.265). If fewer data packets are received, the visual quality of the frame will remain proportional to the amount of received data. This makes Grace’s quality degrade smoother than other loss-resilient coding schemes. Additionally, Grace reduces the ratio of delayed frames up to 4x and the stall rate by 10x which creates an efficient and smooth playback experience.
ADVANTAGES
ADVANTAGES
- Smooth and efficient real-time video communication that is similar in quality to advanced video coding formats
- Utilizes neural networks to decode video frames even with additional packet loss
- Maximizes the user-perceived quality of networked applications
APPLICATIONS
- Real-time video communication (video conferencing, cloud gaming, live broadcasting)
- Autoencoder based video streaming