Jesus Gazol Condon
Jesus Gazol Condon
The training material generation could be focused in the classes that are problematic for the model. This focused training material will be generated as the model is trained, although there...
Test out with real diagram portions. - Make sure that the trained model has all the symbols needed. - Check out the image dimension. Make sure it can detect symbols...
Visualize the training material generated to discover problems. Seems that the problem with symbols out of the visible region could be still there. I need to make sure that the...
The classification examples could be harder, do the following: - Random position of the symbol. - Lines connecting and around the symbols. - Change the font around the symbols.
Check out the default data augmentation in the classifier, the center and crop and other can actually harm the classification if the symbol gets out of the scope.
Have a look at the last layers, check out last section of https://github.com/tensorflow/models/blob/master/research/object_detection/colab_tutorials/inference_tf2_colab.ipynb
- Create a B3 input size backbone with full symbol dictionary. - Check out full process with 10K examples. - Prepare training material for classifier B3 size, approximate 3M -...
The training augmentation uses center and crop, leaving out a bunch of objects out of the picture, remove that augmentation step, hopefully it will be in the configuration.