Video Object Detection at Scale
summary
A video-object pipeline that runs on AWS and Kubernetes handles custom detectors for TV and security use cases, claiming each live stream plus detector needs about an eighth of a GPU per stream for continuous inference; models are ~200 MB and there are millions, so a four-layer cache (S3 → local disk → CPU RAM → GPU memory) and custom orchestration are built to keep hot models in GPU RAM. Detector creation is a few minutes with a UI for quick human-in-the-loop retraining; TPU support is planned.
AI-written summary · about
chapters (10)
- 0:00Introduction to AI talk
- 1:05Raz Zade’s background
- 4:13Matroid’s video object detection
- 11:49Use cases and industries
- 18:31Scaling challenges
- 24:08Custom caching infrastructure
- 29:05Hardware and software evolution
- 39:10Spark and MLlib trade-offs
- 46:15TensorFlow and distributed compute
- 50:54Closing and resources
related talks
14mLive Streaming Object Detection by Leveraging GPU with AVA Platform on EWAOL
35mLarge-Scale Video Intelligence
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
53mVideo Analytics for mere mortals with NVIDIA DeepStream SDK
51mVideo Analytics for Football games
47mBest Practices for Real-time Intelligent Video Analytics
29mImage Processing in the Browser
38mHow We Built a Safety Violations Detection System