How has AI revolutionized geospatial data labeling in less than a month?

AI is experiencing a revival, well maybe coming out of the shadow of crypto & metaverse. The revolution in the world of language models has been visible for several months now. Also, you have probably already seen images generated by AI that could closely resemble real photos. But have you heard about the labelling revolution carried out by the Meta this month?

It is worth starting by reminding how AI models are trained. Neural networks learn by observing. The training data are images marked with labels corresponding to the expected value. The model learns by analysing the given data and what is expected of it. The more labelled data the model sees, the better it is.

In the case of segmentation models, which are the basic class of AI models used in our company, data labelling is very time-consuming, because it requires that each object of the class we want to detect is carefully outlined.?Until now, obtaining such data was done by the tedious manual drawing of masks on all objects.

However, at the beginning of April 2023, Meta published the Segment Anything Model (SAM), which revolutionized our way of thinking about labelling. SAM is an advanced computer vision model that, through the use of machine learning algorithms, can perform image segmentation with very high accuracy. Moreover, it requires very little work from the labeller. The only needed to do is do one of the following:

• clicking in the area of a given object,
• indication of the bounding box in which the searched object is contained,
• indicating several points that belong to the object and some that do not,
• prompt text.

In addition, the SAM is able to operate in real-time, and it is very flexible, thanks to which it can be used in very different areas of computer vision.

For our company, it turned out to be a game-changer. Because of the demand for labelled data, we spend a lot of time labelling every little piece of the curb, every kilometre of asphalt and concrete, and it takes hundreds of hours. Thanks to the use of a tool based on SAM, we can mark 10 times more data at the same time. Such many incoming new data will directly influence the results achieved by the models used by us in the company. The more labelled data, the better models. The better models, the better we are!

Commited by Adam Wisniewski & Jakub ?ukaszewicz