Your Player is Smarter Than You Think: Self-Learning in Adaptive Streaming
summary
Claim: per-title DeepStream delivers CPU-safe base layers and GPU enhancement layers by fine-tuning tiny super‑resolution nets per video, compressed with DeepCABAC and trained only on informative patches via Efficient Patch Sampling (EPS). EPS uses fast DCT complexity scoring to cut training cost up to 83% with negligible quality loss. Playback-side self-training downscales buffered high‑res segments to fine‑tune models on device in ~0.1s, yielding about 1 dB PSNR improvement without extra server retraining.
AI-written summary · about
chapters (8)
- 0:00Self-learning in streaming
- 2:07Video streaming dominates traffic
- 3:53HTTP adaptive streaming
- 5:21Optimizing bit rate ladders
- 8:36Per-title encoding
- 10:00Neural network overfitting
- 13:11Smarter player with self-learning
- 15:40Real-time quality improvement
related talks
47mBest Practices for Real-time Intelligent Video Analytics
17mContent Aware Encoding for low latency live streaming encoders using deep learning
16mSwift: Adaptive Video Streaming with Layered Neural Codecs
30mMLDS 2019: Anubhav Shrivastava of Voot at Machine Learning Developers Summit 2019
21mLearning in situ: a randomized experiment in video streaming
41m8K Codec Low-Latency Livestreaming, and AI Streaming
19mPilotFish: Harvesting Free Cycles of Cloud Gaming with Deep Learning Training
27mEmbedded Deep Learning Super Resolution on GStreamer using ONNX Inference Runtime