Video @Scale 2024 - Day 2
summary
Unified segmentation model Sam 2 extends image-only SAM with streaming memory, object pointers and occlusion scores, and claims 6X faster image inference, 3x fewer video interactions and step‑change accuracy on long videos. A data engine iteratively collects and corrects annotations to build the SAV dataset of ~650,000 challenging masks from 50K+ wild videos. Lighter variants (EfficientSam, SqueezeSam) trade model size—down to ~9.8M params or 63x faster / 32x smaller—for modest accuracy loss, while prefiltering shows 6–15% BD‑rate gains with real‑time runtimes.
AI-written summary · about
chapters (9)
- 0:00Video @Scale 2024 Day 2
- 17:16Day 2 themes
- 19:42Segment Anything Model 2
- 41:08Generative video quality challenges
- 1:02:12On-device segmentation models
- 1:24:55Metrics-driven viewer experience
- 1:43:19Performance signals for recommendations
- 1:57:15Pre-filtering for video compression
- 2:45:06Closing & Q&A
related talks
15mAI-Ready Microservice Architecture
42mSo, You Want to "Do Big Data?"
43mAI/ML Data Pipeline Processing with Go Microservices Based...
08mMaximizing ROI Through Distributed Computing and Multi-Cloud
24mFrom Model to Microservice - Machine Learning at Scale
20mBatching vs. Streaming - Scale & Process Millions of Measurements a Second
45mDeep dive into Unbounded Data Processing Systems
21mScalable agent-based models for optimized policy design: applications to the ec...