Blog

Sep 27, 2023

TinyissimoYOLO AI object detection for low power microcontrollers ...

The team at ETH working on TinyissimoYOLO enabled object detection on industry

The team at ETH working on TinyissimoYOLO enabled object detection on industry microcontrollers with milliwatts of power and with less than 500Kbits of memory for storing convolutional neural network (CNN) weights.

The quantized network architecture has 422 k parameters and enables real-time object detection on embedded microcontrollers, and can use CNN accelerators that are increasingly popular on the chips. In particular, the proposed network has been deployed on the MAX78000 microcontroller achieving high frame-rate of up to 180 fps and an ultra-low energy consumption of only 196 μJ per inference with an inference efficiency of more than 106 MAC/Cycle.

TinyissimoYOLO can be trained for any multi-object detection, but this will increase the size and memory consumption of the network, so the team showed object detection with up to 3 classes with 8-bit quantization on different microcontrollers, such as STM32H7A3, STM32L4R9, Apollo4b and on the MAX78000's CNN accelerator.

The input image size was chosen to support all the common microcontrollers, and the limiting factor is the CNN accelerator of the MAX78000, which does not support CNN inputs greater than 90×91 without using a specialized mode. As a result an input of 88×88 is chosen because it is a tradeoff between maximizing the image size and being able to do pooling on the input dimensions without rounding the dimensions down.

The paper is at: arxiv.org/pdf/2306.00001.pdf

www.eth.ch