Trained model accuracy

[ellaleona]

1. I've trained my own model using roughly 3500 images. I specified 3 classes (Spedies-ID 0, 1, 2). Two are species and one 'empty' category. Upon testing the model, the trained model predicted only two classes (0 or 1), never predicting class 2.

2. I also tried re-training the empty-animal model with this same set of about 3500 training images. I did this after the "empty_animal" model had low accuracy on my 3500 pre-classified images that I tested it on. My re-trained model continued to be very inaccurate (under 50% accuracy).

Do you have any insights or suggestions? Will I simply need plenty of more images to train the model with? Based on issue 1, I'm uncertain that my model is being trained properly, though no apparent errors occur. The model trains in close to 30 seconds, which seems too fast.

My images are atypical for a game-camera; they point down at a platform backdrop rather than out into the field site. I figured this may be why the pre-loaded "empty_animal" model is inaccurate when used on them.

Thanks in advance!
-Ella