How to create robust Classification model?

Why does my classification model fail against unseen real-world testing data, however, the model record validation accuracy of 98% @ Wall time of 31min 22s on the validation set?

P.S. I have used ResNet9, ResNet32 and ResNet152 all fail dramatically against real data (accuracy less than 50%).
Basically, I using cars cartoon images to predict a true car. Since ML is a huge correlation tool I think it is an achievable task? Any thoughts?

The real world data should have some kind of resemblance with your training data, There might be a lot of dissimilarities with your training data and the real world data, my suggestion would be if there are not enough real world data for training include some of the real world data as a validation set and see how the model works.

1 Like