Box label placement

Hello everyone.
I am currently using a Jetson Nano to recognize via camera if a label is correctly placed on top of a box, but things are not going well. Here is sme information about what I’m doing and what’s happening:

  1. I have the jetson with the image which NVIDIA provides with the tutorial. It has some Jupyterlab libraries using pytorch and data entry widgets already created, which I modified slightly to my needs.

  2. The actual approach to the problem is to use image recognition with 3 categories, to determine wheter the label is well placed or not. Categories are: a) well placed, b) wrong corner, c) askew right now.

  3. The images are taken via usb-webcam and manually using the widget. I have a set of 6 boxes with 6 different “labels” (coloured post-its and folded paper sheets with random things written/drawn), which can be swapped beetween them to create 6^6 possible combinations of box/label.

  4. The model in use to process the information is a resNet34.

  5. I have a 7th box/label that are used to test the outputs of the network, and remains outside the training set.

As a final consideration, I have very little experience with Machine Learning (only the tutorial course and a few home made experiments with food recognition so far), which means that I might be doing elementary mistakes. Any suggestion/comment willl be appreciated

Links to some images from the dataset:
https://imgur.com/pB699tk (Askew)
https://imgur.com/ZUtwGsK (Askew)
https://imgur.com/MMqElMs (Askew)
https://imgur.com/dPNnI8j (Well Placed)
https://imgur.com/jJM9h4z (Well Placed)
https://imgur.com/uj8OMWM (Well Placed)
https://imgur.com/8wUTWiQ (Wrong Corner)
https://imgur.com/XTwTZjv (Wrong Corner)