Mismatches

The VisionPro Deep Learning tools indicate a mismatch with an orange marking. A mismatch refers to a feature that is found by the system in an image that you labeled, but there is either a mismatch between a corresponding label, or there are possibly more instances of a labeled feature.

If encountering a mismatch, the first thing to do is to check the accuracy of the labeling on the Training Image Sample Set (among other things like the stability of your lighting and appearance of the feature in the image; see the Image Capture topic for more information).

During application development, a mismatch is indication that the training set is not generalized. For example, applications with very subtle differences between two features could have mismatches because Deep Learning does not have a full understanding of the features yet. In short, there is a difference between what Deep Learning has learned (from the training set) and what it sees in production (processing the test set).

During this phase, you should create a distinct Training and Testing Image Sample Set. It is up to you to manage this, but any images in the training set should be trained (using a 100% fraction).

If you encounter mismatches in your testing set, this indicates that your tool has not been generalized properly. This is generally due to images that don’t properly capture the features, for example there are gaps in your image database. In this scenario, you will need to retrain your tool, making sure that the images/views that contain the mismatches make it into your training set and are properly labeled.