CGI training images

TensorFlow General Discussion
,
1

Hi, I am really new to machine learning and have been playing around with an example from a book. Some models are trained with CGI datasets, and the Rock, Paper, Scissors Dataset works pretty well when I use it to predict real world images.

I tried to replicate training with CGI images using images I rendered myself and I am struggling to understand why my two similar datasets give wildly different results.

Dataset 1 works really well when using real world images, classifying all but one image correctly. But the other Dataset 2 does not work at all, with only one or two images correctly classified.

I randomly split the dataset into 20% validation and 80% training data. The code for training is here.

Is it pure luck that makes Dataset 1 work so well or is there a fundamental flaw in my understanding?

Thank you for advice.

3

Link for dataset 2 is this one.. As I am a new user I can not edit my first post because I can only post 2 links.

4

I suggest you to create more diversity in your dataset e.g.: randomize your background, lighting, position and with other classical augmentations that you can do in 2d or with your 3d engine, shaders etc.

6

If you have an Nvidia GPU available you can also play with NVISII to populate your dataset:

8

Ok thank you for the input. I will give more diversity a try.

I don’t have Nvidia so I will stick with Blender.