My requirement is to create a image classification model for industrial part which has two classes (Valid and invalid). The model recognized the good and fault parts correctly.
But when I provide completely different image for classification (say image of shoes, desk, tree), the model server behaves unpredictably and many times, it classifies the image as good image with 99% probability.
I understand that such images were not part of training in either category, so this is expected. But is it possible to categorize such images in 3rd category (say uncategorized)
I can think of only creating a third class like “Other”. Then add some images with that class, and re-train your model on new data.
Unfortunately, because the third class is everything except the first two, it may not work well for all possible inputs. But if number of possible input images for your domain is limited, then it may actually work.
Thanks for your inputs. I was thinking of similar category: Good, Bad and Uncategorized. But the images in “Uncategorized” class is not limited. It can range to any picture captured randomly.
One related query:
If MxNet model has only 2 categories to classify; and provided image match “Category 1” with confidence of 10%, will MXNet blindly give remaining 90% to “Category 2” (even if “Category 2” does not match)? This is what I am observing in my scenario (weird though), but I am not sure if this is true