Multi-Class Training + Inference

[waschsalz]

Hello.

I was wondering if there currently is an active implementation process for true multi-class training or maybe there is a workaround for that matter.

If not, then I would like to recommend this as a new feature.
This would be a huge improvement since people might want to use just one model for one object and detect anomalies from different angles.

Inference-wise you currently have to use one model for each angle of certain object, which is possible but limiting and definitely more time-consuming. If there is a workaround, please let me know. I am unaware of one hence this feature request.

[alexriedel1]

You could try to train one model with images from all angles. If the variety isn't too large this might work.
Besides there is UniNet which actually is a multi-class model: https://github.com/open-edge-platform/anomalib/tree/main/src/anomalib/models/image/uninet

[waschsalz]

Thank you for your answer. @alexriedel1

I forgot to add in my initial question that I'd only use multi-class capable models for it (UniNet and Dinomaly).

Do you know what steps are typically added / adjusted for multi-class detection compared to single-class workflows? I assume adjusting the threshold is only possible if done manually for each angle.

Directly regarding your answer: Is it as straightforward as adding every OK image from different angles to a train folder?
In the UniNet paper they trained with multiple classes from MVTec (1) and other datasets. In the Dinomaly paper they evaluated MVTec, VisA and Real-IAD under multi-class settings.

[alexriedel1]

Do you know what steps are typically added / adjusted for multi-class detection compared to single-class workflows? I assume adjusting the threshold is only possible if done manually for each angle.

Right, anomalib calculates only one threshold per model. But you could train a model and then use the images of each angle for validation to calculate a threshold for each angle. Then you could build some "sub models" that share the heavy anomaly detection part and only apply different thresholds. You would have to write the code to this, it's not in anomalib

Directly regarding your answer: Is it as straightforward as adding every OK image from different angles to a train folder? In the UniNet paper they trained with multiple classes from MVTec (1) and other datasets. In the Dinomaly paper they evaluated MVTec, VisA and Real-IAD under multi-class settings.

I haven't done this in anomalib yet but I think its the way to go! In the authors UniNet implementation, they simply iterate through all MVTec categories and append them to the training dataset.

[waschsalz]

Right. I will definitely give it a try with my own data and for normal testing purpose with the MVTec categories.

My feature requests stands though because it would be great to play around with multi-class workflows 😄 .

[samet-akcay]

Adding multi-class support is within the scope, but there is no one working on this as of yet