Image Recognition Integration Server IRIS Image Recognition Integration Server
Background
Video Analysis technology
Recent capabilities using deep learning
Speed : up to ~70 frames per second (YOLO-V2 2017) Microsoft ResNet won ILSVRC 2015 with an incredible error rate of 3.6% (Depending on their skill and expertise, humans generally hover around a 5-10% error rate) Speed : up to ~70 frames per second (YOLO-V2 2017) of i From YOLO9000:
Convolution Nets
technologies
Face Detection
Image Captioning
Object Detection
Media Asset Management (MAM)
What is MAM? Software family that: serves as repository of Media Objects - Media files (video, audio, images, captions) - Metadata supports workflow around Media Objects Ingest, Catalog, Transcode, Edit, Distribute, Archive... Large scale (100,000s of large video objects in repository)
Combining MAM and Video Analysis
Problems Video Analysis technology is CPU/RAM/GPU intensive. There are many different models we want to use / rapid evolution.
Need a solution to combine Media Analysis tools with MAM: Scale many parallel jobs. On flexible compute infrastructure (add machines on demand on the cloud). Using many different analysis models and tools (Torch, Tensorflow, Caffe).
Image Recognition Integration Server IRIS Image Recognition Integration Server
Software architecture allowing consumption of: Elastic media analysis engines. Easily extensible in isolated containers for each media analysis engine. Low coupling between MAM and Media Analysis Engines.
IRIS Detection Server Client Dalet API Task Manger Worker Queues REST request API call Task Manger Queues Worker REST requests Celery TASK Celery TASK API call
Support Various detection technologies
IRIS Detection Server Task Manger Worker Queues REST requests Celery TASK Celery TASK API call
IRIS Detection Server Task Manger Worker Worker Queues REST requests Celery TASK Celery TASK API call
Scalability and Stability
IRIS Detection Server Queues Task Manger Worker Worker Worker Worker REST requests Celery TASK Celery TASK
IRIS Detection Server Task Manger Task Manger Task Manger Queues Worker Worker Worker Worker Celery TASK Celery TASK REST requests
Friendly Interface
Technology Used
Result
Detection server Manger- RESTful API Yolo Worker Manger- celery
Video
Manger got the job and sended task to queue Job received from client 2 Manger got the job and sended task to queue 1 Job received from client 3 Worker start task
6 5 4 Job received the results Manger handle the results and pass to the Detection server 4 Task finished