1. Get a lot of data e.g Imagenet
2. Create Function for online(during training) generating Data
3. Use some CNN as Feature extractor
4. Train classifier
Classify as 0 - no angle rotate
1 - image rotated
Classifier with 360 labels does not work
First test model was trained with ResNet-18 and got:
* HIGHREST ACCURACY: 0.7892497518082542
Repo uses resources from:
https://github.com/DFGANDP/Pytorch-EarlyStopping
- Train on more Data
- Use more complex models (e.g. Vision Transformers)
- Write Real Time solver With Selenium?