In Part #5, the model performs poorly on my own set of images. As the last part suggests the error might be related to some distributions and asks to come up with potential solutions. Does that imply the solution is using our set of photos to train the network or some form of data augmentation so the NN learns that different lighting etc. doesn’t change the labelled output? I tried using DropOut too but the performance was about the same as without which hints at an issue with the training? Curious to know what the other solutions are
Hi @Wilbert,
This is an open ended question, having different answers. You can try whatever comes to your mind and see which ones get you the better results.
Best,
Mubsi
2 Likes