Check out what’s in store for SIGGRAPH 2020, 19-23 July 2020 in Washington, D.C.!

Learn More
Exhibitor Sessions: Intel Exhibitor Session: Build a Deep Learning Video Analytics Framework for Intel AI Platforms
Event TypeExhibitor Sessions
Interest Areas
Arts & Design
Gaming & Interactive
New Technologies
Production & Animation
Research & Education
Primary Interest Areas
New Technologies
Registration Levels
EX
XP
F
FP
S
B
E
TimeTuesday, 30 July 20192:30pm - 3:30pm
LocationRoom 406B
DescriptionAs internet traffic is dominated by video content, Intel chief architect Raja Koduri said image/video is the new data type as majority of the data for AI computing are images/videos.

Traditional video usages are extending to AI based video analytics, deep learning has been used in various segmentations in IOTG and DCG. The top usages are image detection and classification for digital surveillance, manufacture defect detection, video content classification and summarization, target Ad insertion, content compliance and content monetization, etc.
Intel has been developing many different AI hardware's, including CPU, GPU, VPU, FPGA and ICE, this creates a potential problem for customer SW to support video analytics for various platforms and HW IPs.

This presentation will describe how do we build a unified framework based on FFMPEG, GStreamer to enable a video analytics on all Intel HW.

1. We will go over the overall video analytics architecture, including media, openVino, mkl-dnn, cl-dnn and OpenCL along with key usage data flow.
2. We will focus on one framework and one API concept, how do we design a common meta data format to be shared between FFMPEG and GStreamer, how do we build FFMPEG filters and GStreamer plugin for video analytics inference stack.
3. We will then deep dive into media and compute interoperability for buffer sharing, synchronization, we will describe Inference tensor data type and its limitation and how to add media data type into inference path natively.
4. We will further describe a workload scheduler that allows scheduling cross various IPs in the same PCI device, and cross multiple PCI devices, and cross cloud/edge.
5. At the end, we will further discuss four key performance indicators ((throughput, efficiency, latency and accuracy), how we design a framework that improve the KPIs and beat competition.

Speaker: Charlie Wang, Intel

Contact Us