Best Practices for Real-time Intelligent Video Analytics • Ekaterina Sirazitdinova • GOTO 2021

46:32 543 views 100% Published 3 months ago

This presentation was recorded at GOTO Copenhagen 2021. #GOTOcon #GOTOcph

Ekaterina Sirazitdinova - Data Scientist for Computer Vision, Video Analytics & Deep Learning at NVIDIA

Larger and more complex Vision AI networks enable better accuracy and precision since they are able to encode more information. This increase in size and complexity is, in turn, naturally associated with trained AI models having lower throughput and larger memory requirements. With that, real-time AI inference is becoming a new great challenge in intelligent video analytics.
NVIDIA’s approach at solving this problem relies on two major components: first, tuning AI models for performance depending on the target deployment hardware platform, and, second, optimizing the use of available GPUs.
From this talk, you will learn how to leverage this approach to achieve real-time inference performance by using software tools like DeepStream SDK, TensorRT and Triton Inference Server [...]

00:00 Intro
01:23 Why intelligent video analytics?
06:27 Challenges with intelligent video analytics
08:49 Tips & tricks for efficient AI video analytics
09:00 Transfer learning
11:03 Data augmentation
12:19 Automatic mixed precision (AMP)
14:03 Quantization
16:18 Network pruning
18:10 Network graph optimizations
19:51 Kernel auto-tuning
20:52 Dynamic tensor memory upon inference
21:33 Multistream concurent execution
22:04 Free NVIDIA products designed to make your AI App efficient
23:10 NVIDIA's end-to-end AI workflow
25:32 TAO toolkit
28:45 high performance pre-trained vision AI models
30:02 Enabling beyond pre-trained AI models
30:48 Achieving state of the art accuracy for public datasets
31:10 NVIDIA TensorRT
32:48 TensorRT workflow
33:56 Triton inference server
37:12 DeepStream SDK
38:57 DeepStream application architecture
41:23 Pipelin efficiency with zero memory copies
42:19 NVIDIA graph composer
42:32 DeepStream video demo
43:54 Summary
45:18 Developer resources
46:09 Outro

Download slides and read the full abstract here:

Magnus Ekman • Learning Deep Learning •
Phil Winder • Reinforcement Learning •
Kelleher & Tierney • Data Science (The MIT Press Essential Knowledge series) •
Lakshmanan, Robinson & Munn • Machine Learning Design Patterns •
Lakshmanan, Görner & Gillard • Practical Machine Learning for Computer Vision •
Aurélien Géron • Hands-On Machine Learning with Scikit-Learn, Keras & TensorFlow •
#NVIDIA #AI #ML #DataScience #AINetworks #ArtificialIntelligence #MachineLearning #DeepStream #DeepStreamSDK #TensorRT #NVIDIATensorRT #Triton #TritonInferenceServer #AIAnalytics #AIVideo #AIVideoAnalytics #DataAugmentation #AMP #AutomaticMixedPrecision

Looking for a unique learning experience?
Attend the next GOTO conference near you! Get your ticket at
Sign up for updates and specials at

SUBSCRIBE TO OUR CHANNEL - new videos posted almost daily.

Watch on YouTube