Looking for any suggestion. I have a model that is trained to classify 2 different types of animal. It seems to work fine when one of them is detected but if any other animal is detected it still gives a result of one of them.
Is there a recommended way to have a catchall that will classify the other type as not known or something?
Hi @delfin4,
Neural networks always have to predict an output. We usually recommend having a “random” class with pictures of random objects/animals that are not the 2 you want to classify. This will make the model more robust.
Aurelien
I did wonder about that. Any suggestions where a decent set of random images can be downloaded? This model will be deployed as part of a conservation effort in Africa so city images, etc aren’t much use:)
Another question - Is it worth going through and labeling each image with bounding boxes or just use the whole images for classification?
And another question - why would this image get classified as an African Elephant and not be “uncertain”
Hello @delfin4,
Is it worth going through and labeling each image with bounding boxes or just use the whole images for classification?
That depends really on the use case, if you don’t need to location of the object in the image, nor the possibility to count the objects, use an image classification model, it will be smaller
why would this image get classified as an African Elephant and not be “uncertain”
You can set the confidence threshold in the “Model Testing page”, by clicking on the three vertical dots next to classify all:
If the value is below the confidence threshold, it would get classified as uncertain.
Best regards,
Louis