Custom vision-based AI software created on Nvidia’s Tao tool set can be run on Arm based Cortex-M, Cortex-A and Ethos-U microcontrollers according to Edge Impulse, which teamed up with Nvidia to create the work-flow.

Nvidia Tao visual inspection in Edge Impulse
“Engineers can create computer vision models that can be deployed to hardware including NXP I.MXRT1170, Alif E3, STMicro STM32H747AI and Renesas CK-RA8D1,” according to Edge Impulse. “The platform allows users to provide their own custom data with GPU-trained Nvidia Tao models like Yolo and RetinaNet, optimising them for deployment MCUs, MPUs and accelerators.”

The model list is: RetinaNet, YOLOv3, YOLOv4 and SSD object detection, and image classification.

To the user, Tao’s options appear inside Edge Impulse – which is the name of its tool as well as the company. “Engineers can finally utilise Nvidia’s AI models on hardware outside of that offered by Nvidia, a capability exclusively provided by Edge Impulse,” it said. “Take your own models or pre-trained models, adapt them to your own real or synthetic data, then optimise for inference throughput. All without needing AI expertise or large training datasets.”

Over 100 example computer vision models are accessible via the combined work flow: 88 object detection architectures and 15 image classification architectures.

These include pre-trained 3x224x224 backbones from Nvidia’s Tao catalog and others trained by Edge Impulse on ImageNet. For image classification, pre-trained weights only support images between 32 x 32 and 224 x 224.