Region-based Content Enhancement for Efficient Video Analytics at the Edge
summary
RegenHance presents a concrete system that enhances only important frame regions to cut compute and bandwidth: a macroblock-based region importance predictor runs ~200× lighter than pixel-level detection, a region-aware enhancer stitches sparse regions into dense tensors for efficient SR (using a Tetris-style bin packer), and a profile-based execution planer allocates CPU/GPU work. Prototype on five heterogeneous edge devices evaluates object detection and semantic segmentation and shows region-based enhancement raises accuracy by 10–19% and boosts throughput 2–3× versus frame-based SR.
AI-written summary · about
chapters (10)
- 0:00Region-based video analytics
- 1:19Bandwidth and accuracy challenges
- 2:42Limitations of super-resolution
- 3:46Region-based enhancement concept
- 4:52Region importance prediction
- 6:39Region allocation across streams
- 8:27Tetris bin packing for stitching
- 10:37Resource allocation
- 12:00Performance benchmarks
- 13:03Conclusion
related talks
22mAI Enhanced GPU Video Coding: Achieving Joint High Compression Efficiency and Throughput
33mArchitecting Scalable Multi-Modal Video Search
01h 56mEdge AI Systems
17mYuZu: Neural-Enhanced Volumetric Video Streaming
25mOptimal Video Compression Using Pixel Shift Tracking
35mLarge-Scale Video Intelligence
16mGemel: Model Merging for Memory-Efficient, Real-Time Video Analytics at the Edge
11mSaFace: Towards Scenario-aware Face Recognition via Edge Computing System