FOMO, Faster Object Detection On Tiny Processors
Edge Impule recently announced FOMO, a powerful but small-scale object detection model geared to run on low-power processors at high speeds.
Created on April 18|Last edited on April 18
Comment
Recently, Edge Impulse announced a new ML model they have developed called FOMO (Faster Objects, More Objects). This model is designed for faster image detection than other image detection models, like YOLOv5.
FOMO's goal is to be able to be run on low-end devices, like small raspberry pi's or other small device processors, while maintaining high speeds and accuracy with a few drawbacks.
How does FOMO compare to other object detection models?
First, it's good to know the difference between image classification and object detection. Image classification takes an image or video input and tries to identify what it represents as a whole, based on what it knows like telling what animal is in the picture. With object detection, the goal is to identify individual objects withing an image or video, generally keeping track of what the objects are, where they are, and how big they are in the view.
FOMO is designed as an object detection model able to quickly identify many objects in a given image or video stream while being run on extreme low-end hardware. Unlike it's counterparts, FOMO loses the ability to identify the size of objects within the image in favor of speeding up and sizing down the model. Because of this, FOMO is best used when objects within the image are of similar size.
How does FOMO work?
Because it's only looking for object type an position, a lot of the precision can be cut out from the process, one of the big things being that FOMO ditches the traditional bounding boxes other models use for judging object size.
FOMO works at a variety of input resolutions. In the process, it splits the input image into a grid of 8x8 pixel (by default) cells and effectively runs image classification on it, running it through convulutional layers, eventually resulting in a heatmap of what objects it thinks are where in the image.
This stripped down and speed-optimized process allows FOMO to run on low-power processor setups easily while maintaining usable speeds.
Find out more
Add a comment
Tags: ML News
Iterate on AI agents and models faster. Try Weights & Biases today.