For those building AI systems in manufacturing: how are you approaching vision-based anomaly detection today? Are you using off-the-shelf solutions/models, fine-tuning foundation models, or building models from scratch? What were the key trade-offs that influenced your choice, and how has it performed in production? Curious to learn from real-world experiences. Thanks!

What kind of anomaly are you talking about. Depends on what are you trying to detect. We have a manufacturing client where we need to separate beans based on size, texture color etc and divide it into different categories based on it. Approach without vision: Take the reading of the beans and do a regression on the data. With vision: Fine tune a model with custom data and check the result. If unsatified you will have to add more parameters/imporove -dataset etc

Just to echo what Cookie said: I'm sure it depends on what the data / anomaly's look like. I've had good success in the past fine-tuning off the shelf models with data from the floor. Both for object detection or classification use cases. Are you allowed to share a bit more on what the data looks like?