Large-Scale Video Intelligence
summary
Claims a scalable multi‑modal video search by embedding detected entities rather than whole videos and demonstrates it works at scale. Batch pipeline uses YOLO v11 + BoSort for detection/tracking, key‑frame deduping, SigClip 2 fine‑tuned plus ArcFace for faces, stored in Quadrant vector DB. GPU optimizations use NV video codec decoding, zero‑copy via DLPack, CUDA ops and TensorRT inference, yielding ~10x throughput gains and about 20x real‑time on an A100; an on‑demand agent (Qwen family visual LLMs) handles deep analysis.
AI-written summary · about
chapters (6)
- 0:00Video intelligence project
- 3:22Decarb video search system
- 15:15Pipeline efficiency
- 21:00Implementation lessons
- 24:35System limitations
- 26:27Q&A
related talks
33mArchitecting Scalable Multi-Modal Video Search
53mVideo Object Detection at Scale
18mCoVA: Exploiting Compressed-Domain Analysis to Accelerate Video Analytics
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
46mML Use Cases at Think Big Analytics
24mBeyond the Frame: Using Computer Vision to Unlock the Potential of Video Data
47mBest Practices for Real-time Intelligent Video Analytics
28mVision, Big Data and AI