Video Analytics with Zero-streaming Cameras
summary
DIVA is a camera/cloud runtime that leaves video on local SD cards and answers queries by (1) building sparse landmark frames (e.g., one every ~30s) that bootstrap tiny on-camera neural operators, and (2) running multi-pass operator upgrade during a query so cheap rankers quickly surface likely frames and stronger models refine results. Evaluated over 15 videos (720 hours), DIVA runs at >100× real-time, cuts end-to-end latency 3–4× versus baselines, and yields >1000× network reduction.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:18Trends of surveillance cameras
- 2:14Zero-streaming (ZS) cameras
- 3:16DIVA: a runtime for ZS cameras
- 3:54Supported query types in DIVA
- 5:17Basics: operators
- 6:20Design #1: landmarks (capture time)
- 7:41Design #2: operator upgrade (query time)
- 9:03Concrete execution plan
- 10:20More about DIVA
- 11:09Experiment settings
- 11:40Highlighted results
- 13:40Summary
related talks
47mBest Practices for Real-time Intelligent Video Analytics
23mAVA: Towards Agentic Video Analytics with Vision Language Models
28mVision, Big Data and AI
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
17mRECL: Responsive Resource-Efficient Continuous Learning for Video Analytics
15mBoggart: Towards General-Purpose Acceleration of Retrospective Video Analytics
46mComputer Vision to Secure Your Surroundings with AI/ML Solution Built Using Open Source Tools at the Edge
38mHow to Build Video Analytics for Live Streams Without Big Tech Resources