Swift: Adaptive Video Streaming with Layered Neural Codecs
Talkat 19th USENIX Symposium on Networked Systems Design and Implementation (NSDI '22) · 2022-04-04..04-06·
summary
Swift implements a layered neural codec that encodes frames as a base code plus residual layers, a single-shot decoder that fuses any subset of layers in one pass (latency independent of layer count), and an ABR policy that trades off network and client GPU compute via multi-exit decoding. It claims 3x faster reaction to throughput jumps, 45% higher QoE versus the next best, and up to 18% bandwidth savings while running on commodity GPUs/accelerators.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:35Outline
- 1:14Video Streaming Today
- 2:46Limitations of Today's Streaming Alg
- 4:54Existing Alternative: Layered Video St
- 6:37Swift: Layered Coding Using Neural
- 7:472 Layered Neural Video Decod
- 10:463 Streaming Protocol/Algorithm
- 11:57Putting Everything Together: Sv
- 12:28Swift's Performance Comparist
- 15:04Summary
related talks
17mLiFteR: Unleash Learned Codecs in Video Streaming with Loose Frame Referencing
16mGRACE: Loss-Resilient Real-Time Video through Neural Codecs
18mOn-device Video Playback Upsampling
22mProviding better video experiences for the next billion users.
17mYour Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
18mScaling AV1 End-to-End Delivery at Meta
17mContent Aware Encoding for low latency live streaming encoders using deep learning
31mAV1 Update