Best Practices for Real-time Intelligent Video Analytics
summary
NVIDIA workflow for real-time video analytics combines DeepStream SDK, TensorRT, Triton Inference Server and the TAO toolkit to go from pre-trained models to deployed pipelines. Techniques include transfer learning, online data augmentation, AMP (1.5× faster training), INT8 quantization, pruning (up to ~10× smaller models reported), layer fusion, kernel auto-tuning, dynamic tensor memory and multi-stream execution. DeepStream runs a zero-copy GPU pipeline, supports batching and DeepSort tracking, and claims up to ~10× inference speedups after TensorRT optimization.
AI-written summary · about
chapters (30)
- 0:00Intro
- 1:23Why intelligent video analytics?
- 6:27Challenges with intelligent video analytics
- 8:49Tips & tricks for efficient AI video analytics
- 9:00Transfer learning
- 11:03Data augmentation
- 12:19Automatic mixed precision (AMP)
- 14:03Quantization
- 16:18Network pruning
- 18:10Network graph optimizations
- 19:51Kernel auto-tuning
- 20:52Dynamic tensor memory upon inference
- 21:33Multistream concurent execution
- 22:04Free NVIDIA products designed to make your AI App efficient
- 23:10NVIDIA's end-to-end AI workflow
- 25:32TAO toolkit
- 28:45high performance pre-trained vision AI models
- 30:02Enabling beyond pre-trained AI models
- 30:48Achieving state of the art accuracy for public datasets
- 31:10NVIDIA TensorRT
- 32:48TensorRT workflow
- 33:56Triton inference server
- 37:12DeepStream SDK
- 38:57DeepStream application architecture
- 41:23Pipelin efficiency with zero memory copies
- 42:19NVIDIA graph composer
- 42:32DeepStream video demo
- 43:54Summary
- 45:18Developer resources
- 46:09Outro
related talks
53mVideo Analytics for mere mortals with NVIDIA DeepStream SDK
27mOn challenges and considerations for real time AI processing
36mVideo analytics at scale: Deep Learning, Computer Vision, ML & Spark
18mDeploying ML solutions with low latency in Python
45mExtracting real-time data at the edge with SageMaker & NVIDIA GPUs
54mDeep learning on AWS with NVIDIA: From training to deployment
20mRealtime Video AI with Diffusion Models
35mLarge-Scale Video Intelligence