Best Practices for Real-time Intelligent Video Analytics

Explore best practices for real-time intelligent video analytics in this comprehensive conference talk from GOTO Copenhagen 2021. Dive into the challenges of implementing Vision AI networks and learn NVIDIA's approach to achieving real-time inference performance. Discover techniques such as transfer learning, data augmentation, automatic mixed precision, quantization, and network pruning. Gain insights into NVIDIA's end-to-end AI workflow, including the TAO toolkit, TensorRT, Triton Inference Server, and DeepStream SDK. Learn how to optimize AI models for various hardware platforms and maximize GPU utilization. The presentation covers practical tips, free NVIDIA products, and demonstrates real-world applications through a video demo. Ideal for data scientists, AI developers, and anyone interested in cutting-edge video analytics technology.

Intro
Why intelligent video analytics?
Challenges with intelligent video analytics
Tips & tricks for efficient AI video analytics
Transfer learning
Data augmentation
Automatic mixed precision AMP
Quantization
Network pruning
Network graph optimizations
Kernel auto-tuning
Dynamic tensor memory upon inference
Multistream concurent execution
Free NVIDIA products designed to make your AI App efficient
NVIDIA's end-to-end AI workflow
TAO toolkit
high performance pre-trained vision AI models
Enabling beyond pre-trained AI models
Achieving state of the art accuracy for public datasets
NVIDIA TensorRT
TensorRT workflow
Triton inference server
DeepStream SDK
DeepStream application architecture
Pipelin efficiency with zero memory copies
NVIDIA graph composer
DeepStream video demo
Summary
Developer resources
Outro