Bringing AI Vision Capabilities
to
Tiny Devices!
Object Recognition using Instance Segmentation on low-power MCU
Unlock affordable, accurate, object detection, identification and recognition on low-powered edge devices with embedUR. See a state-of-the-art Instance Segmentation showcase using YOLO on a battery-powered MCU (with NPU accelerator) to enable a wide range of AI vision applications.
Object Recognition using Instance Segmentation on low-power MCU
Unlock affordable, accurate, object detection, identification and recognition on low-powered edge devices with embedUR. See a state-of-the-art Instance Segmentation showcase using YOLO on a battery-powered MCU (with NPU accelerator) to enable a wide range of AI vision applications.
As embedded systems experts, we frequently collaborate with silicon vendors to put their latest chips through their paces, optimize the firmware and build SDKs so others can use the chips in their apps, without needing to know the low-level instruction set.
As we get deeper into deploying Al and ML on edge devices, this role has extended to include the
creation of basic training models for high-level functions. This project is an example of this, in which
we were given a chip, and not much else.
Image Segmentation Comparison
This video compares image segmentation running on a cloud server with GPUs, a laptop with a GPU, and Synaptics’ battery-powered AI-Native IoT platform (SR110) designed for tiny devices.
It is pretty astonishing that the tiny SR 110 is able to keep up with and performs with the same functionality as a laptop and cloud server which cost considerably more and consume tons of power.
Image Segmentation Comparison
This video compares image segmentation running on a cloud server with GPUs, a laptop with a GPU, and Synaptics’ battery-powered AI-Native IoT platform (SR110) designed for tiny devices.
It is pretty astonishing that the tiny SR 110 is able to keep up with and performs with the same functionality as a laptop and cloud server which cost considerably more and consume tons of power.
Figure it out, they said. So we did!
Figure it out,they said.
So we did!
Solution Components
Even the creators of the chip were blown away by our accomplishments!
How was this built?
In order to bring fast Object Identification image processing to life, on a tiny module, we needed to
adapt the YOLACT Instance Segmentation model from Ultraytics, to run on Synaptics’ latest Astra MCU.
We used Tensorflow Lite to achieve this. Then the model was trained using the COCO dataset.
How was this built?
In order to bring fast Object Identification image processing to life, on a tiny module, we needed to
adapt the YOLACT Instance Segmentation model from Ultraytics, to run on Synaptics’ latest Astra MCU.
We used Tensorflow Lite to achieve this. Then the model was trained using the COCO dataset.
Partner with embedUR
Got an innovative product idea you need help getting started? We’ll find the right way to bring it to life. embedUR is your go-to embedded partner for turning smart device ideas into reality. With 20 years of embedded engineering experience at the edge, we bridge the gap between the latest technologies, frameworks and tools and your vision.
Partner with embedUR
Got an innovative product idea you need help getting started? We’ll find the right way to bring it to life.
embedUR is your go-to embedded partner for turning smart device ideas into reality. With 20 years of
embedded engineering experience at the edge, we bridge the gap between the latest technologies,
frameworks and tools and your vision.