Newbie looking for directions

[tomy983]

Hi all, I wanted to ask for general direction on how could anomalib be used in a production line.
I have been playing with yolo and started to understand how that works, I also did all the anomalib notebooks I cold find (had some issues running some of them). Now I am wondering: let's say I fix a camera on a production line (like the one shown in the autodistill library here) could you describe me what the steps would be to use anomalib to detect for example an half empty bottle, a green cap or an unknown issue with the bottle?

If I am understanding all this correctly the steps could be:
1.Collect dataset (picture, video)
2.Label (pictures or frames)
3.Train for object detection yolo or similar
4.Run inference with object detection and somehow use the output to train anomalib one picture/frame at the time (what if the frame only has half bottle?) Suold I be using segmentation masks for this process?
5.From live video use frames for inference first with object detection and use the output as input in anomalib.

Do the steps above make any sense? Is there another better or more suitable approach?
Is there a notebook or some project using a real world example for quality control?

Sorry for the basic questions, really new to all this.

[tomy983]

Maybe my question is not relevant/properly formed.
I did see and looked for the meaning of "unsupervised" while looking at the documentation of a Anomalib.
If I undestand this correctly, this means I should not need any lables in the dataset. But while this concept I can see how might be suited for something like the MvTec dataset with still pictures of an object, how can all this work for a video? Is it even possible? Can the dataset be something like this as it is (a video of boxes moving on a conveyor belt)?

Or does the video need to be recorded/processed somehow(object detection-segmentation-only certain frames passed for training/inference)?
Are there any examples of anomalib working with video feeds for manufacturing? (I saw the robot notebook, but looks like it is still using pictures instead of videos...)