If you’ve been keeping up with us for a while, you won’t be surprised to learn that we’re always on the lookout for new technologies. We’re especially curious on how they perform in a manufacturing context: showing flashy results is one thing, but using AI in real-life industrial settings is another thing entirely.

When Meta released the Segment Anything Model (SAM), it immediately caught the attention of our AI experts. Our E&I team spends a lot of time on manually identifying and segmenting the components of electrical cabinets, so they wondered if this new model could help us out.

Easier said than done, because traditional computer vision techniques often struggle with the details and variations found in these cabinets. SAM, however, offered a promising solution - especially if we took some time to tweak it to our needs.

In this whitepaper, Arno Van Eetvelde and Nick Remouchamps explain how they used Low-Rank Adaptation (LoRA), a relatively small dataset, and a custom solution to train SAM on the specific nuances of electrical components.