Observe the world through the eyes of a computer.

With the web offering a nearly unlimited amount of images and videos, it’s incredibly easy to start utilising these sources to train computer vision models. Pre-training means the hard work of gathering this data and structuring it in a way that allows you to start training has already been done, meaning you can get up and running with computer vision really quickly.

The technical nitty-gritty

A pre-trained model has been trained on a publicly available dataset. Most models make use of ImageNet, an image database that is organised in a way that every entity that is recognises in the database is depicted by thousands of images, from every angle, direction and light level. This means that it can be recognised in a variety of different contexts. Then, depending on the model architecture, the AI model reads the images pixel-by-pixel and tries to learn how to recognise patterns that are observable in all images that are related to the main entity, and, most importantly, what makes those objects unique as compared to other entities in the dataset. Put simply, this means that it will automatically learn that apples are rounder than pears, and therefore when estimating what kind of fruit a vegetable is, the values of the pixels that form the edges of the shape are seen as more important than the corner pixels – they don’t hold any significant value to determine whether an image represents an apple or a pear.

What is our opinion?

If pre-training is an option for your use case, we will always recommend it as an option to try before diving into custom training. We’re a huge fan of pre-trained AI models because it allows you to start with AI right away, without having to wait on data collection. This saves you time (and money!) because it allows us to largely skip the long process of data cleaning and pre-processing. Especially for (custom) image or video processing, there’s often a need for very large datasets when you don’t use pre-trained models.

How can you apply it?

Pre-trained computer vision models often at times allow you to predict a set amount of classes. This means from the get-go, you are limited in the applications for your model. However, this does not mean the end! There are literally a thousand classes that some models support, ranging from a goldfish to an oil filter. And if that doesn’t suit your case, these models allow you to export a frame of an image that contains a more general class, which you can then use to do some more customisation with traditional methods. For example, you could export a more generic class “bottle” to find all bottle shapes, and then take all the instances of bottles and compare them to more specific whisky bottles to locate the exact brand that you’re observing in your image.