Senticore presence in the UK
Cloudwise: About us
We’re a consultancy with a proven track record in applying open-source technology to solve business problems in the big-data & ML domain.
Recent projects have focussed on computer vision-based applications using AWS, Python, OpenCV and TensorFlow Lite at our consulting company, cloudwise.
Prior to this, I built, led and mentored a cross-regional team at Cisco that conceived, implemented and delivered the Linux Foundation PNDA.io data platform project to network service providers.
I’m also passionate about mentoring and developing individuals and I have extensive experience of servant leadership in agile-focused, R&D and creative teams.
Technology & Innovation Consulting Machine Vision Engineering
1. Visual surveillance
2. TensorFlow Lite 2 on armv6
3. Region-mapped CNNs
Customer: visual surveillance
1. Customer problem statement
• Deliver a reliable person detection solution from legacy, low-resolution (VGA) RTSP-based security cameras
• Utilised pre-trained SSD-MobileNet v2 & YOLO v3 CNN models in the OpenCV v4 framework using Python 3
• Per-frame inference (SSD-MobileNet) @ ~500ms (AWS EC2 micro instance)
• Deployed and running on t2.micro AWS-Linux-based EC2 instances
• 2 concurrent VGA camera streams @ 30 fps per instance
• FTP image uploader on presence detection
• Code and utilities committed to open-source at github
Open-source : TensorFlow Lite 2 on armv6
1. Problem statement
• Investigate the viability of running presence detection at the edge on armv6-based
• Extend the ‘hello_world’ code in the TF Lite 2 repo to support object recognition as well as object detection
• Build script support for armv6 target
• Extended C++ image_label app to support object recognition via SSD MobileNet v1
pre-trained, quantised model
• Benchmarked basic object recognition on RPi armv6 and armv7 platforms
• Code committed to open-source at https://github.com/cloudwiser/ObjectDetectionRPiZero and https://github.com/cloudwiser/ObjectDetectionRPi (armv7)
Research: region-mapped CNNs
1. Problem statement
• For fixed viewpoint image streams, enable separate regions of the frame to be parsed by different CNN models
• Enables ‘best-fit’ selection of the CNN model for each region based on performance, accuracy (mAP) and functional
• For example
• Model A = lower accuracy, higher-throughput CNN for peripheral viewport regions
• Model B = higher accuracy, lower-throughput CNN for central viewport region-of-interest
• Frame is divided into a m x n grid depending on viewpoint
and detection requirements
• Image regions are re-shaped to fit input image resolution for
the respective model
• Regions are overlapped (padded) to reduce the probability
of missed detections on the intersections
• Anchor boxes are also re-shaped based on region shape and
objects of interest
Customers & Partners
Senticore understand databases very well in general, and has a wide network of connections to various technology experts, particularly in Israel. Senticore has a very good working relationship with IBM and would recommend them.
Dr. Andreas Zekl
We collaborated with Senticore for many years, including participating in projects involving IBM research and development and consulting for strategic customers. Senticore is a database expert and knows a lot of people-both inside IBM expert community and beyond it, especially for databases and enterprise applications.
Senticore has been supporting Husqvarna for more than 10 years. They have been assisting Husqvarna in designing and implementing Multi-site systems and workflow for PDM-solutions. Senticore’s focus is on a distributed environment, performance and general database issues.