I'm encountering what I believe can only be deemed a bug with the deep learning tool "label objects for deep learning."
In summary, changing the class value of a training sample in the manager does not actually change its class value.
I prepared a feature layer of training samples in hopes of automating parking lot digitization on our campus. This feature layer is composed of 4 different types of features, or "classes," blacktop, sidewalk, curbs, and marking lines.
The schema is simple. It's the 4 classes, blacktop, curbs, sidewalks, marking lines, with values of 1, 2, 3, and 4, respectively.
So, I import the feature layer, then assign each sample to match its respective class identified in the schema. This appears to work, but it does not actually, and this is where the bug comes in. Though it appears as though its been updated in the training samples manager, it doesn't actually update the internal class value for a training sample.
For example, lets say I'm updating a sidewalk feature because it imported as class_497 automatically. I update it to sidewalk, class value 3, and it appears as though it worked. I save the training samples and close Pro for the day. The next day, I come back in and reopen the tool. I import my data that I'd saved the previous day and that sidewalk sample is class_497 again, not my sidewalk class I specified yesterday. The value never actually updates internally. So, while the training samples manager can make it appear as though the class value is actually updated when fixing up your samples, its not.
I've attached a video of the workflow so you can see exactly what I mean. Updates made to the training samples are purely cosmetic, as the underlying data does not change.
Please let me know if I'm missing something! So far, the only way I can get a sample with the proper class value is by making a new feature within the sample manager itself.
