About generated symbols

[youmizai]

Thank you very much for the source code provided, I am trying to use your data generation algorithm. However, the generated data is highly distorted and does not improve the recognition rate. According to your paper, the distortion factor is set to 0.3.

[kdavila]

Hello. Are you using the CROHME competition datasets?

[youmizai]

Yes, I am using the CROHME2016 dataset, the number of symbol categories is not balanced in the dataset.

[kdavila]

Exactly. On my original paper that is also part of our conclusion. There is some benefit to add the synthetic data at the per-class average accuracy, but due to the huge difference in class representation, adding extra synthetic samples can increase the confusions between classes with very similar shapes, which are very hard to distinguish in isolation. In other words, as the recall for some classes such as "\times" increases, the recall for other classes such as "x" decreases, but "x" is way more common in the test set than "\times". I would suggest comparing results with and without the synthetic data at the confusion matrix level, not just based on global accuracy.

[youmizai]

Ok, thank you very much for your suggestion.