Organizers:

Alexander Bovyrin (Algorithm Innovation Lead, Sr. PE, Intel)
Yury Gorbachev (OpenVINO Lead Architect, PE, Intel)
Pavel Druzhkov (Sr. Machine Learning Engineer, Intel)
Nikita Manovich (Sr. Software Engineer, Intel)

Time: 13:30-17:30 (Half Day — Afternoon)

Description:

This tutorial aimed to help deep learning/computer vision engineers productize their solutions bringing them to the edge by using Intel® OpenVINO™ tools. Deep Learning based algorithms are very resource demanding and compute intensive tasks. Selection of deployment tools is important choice that has to be made whenever algorithm is ready for production. In practice, selection of tools not only impacts speed, but also quality of the final solution. Hence, knowledge about target platform and features that are supported by enabling tools at algorithm selection stage can be critical.

We cover following important topics:

• Introduction to OpenVINO toolkit:
  • Architecture of Deep Learning deployment tools;
  • Cross-platform portability for DL deployment;
• Most important features:
  • Topology specific optimizations.
  • Deep network graph optimization (hardware agnostic and hardware specific) with samples.
  • Heterogeneous execution. Sample.
  • Asynchronous execution. Sample.
  • Various precision support. Sample.
  • Dynamic batching. Sample.
• OpenVINO Model Zoo:
  • Models overview and performance characteristics.
• OpenCV DNN module:
  • DNN model overview.
  • Samples of how to use OpenCV DNN.
  • OpenCV efficiency with Inference Engine backend.
• Helper tools and planned features:
  • CNN optimization tools: quantization, pruning, sparsity.
  • CVAT – open computer vision annotation tool.

Tutorial will also walk participants through actual examples of algorithm design (topology selection and quality evaluation) and workflow of porting to Intel DL Deployment tools. Performance figures for final solution as well as comparison with other Deep Learning solutions will be provided.

A G E N D A