Data augmentation for dense correspondence?

Question

I build a dataset by CAD software for dense correspondence task in computer vision. I trained on that data but got a severe overfitting. Then I used data augmentation, while the loss on validation set didn't decrease. I even gradually increase the extent of data augmentation during the training, in other words, at first there is nearly no data augmentation, and the extent of augmentation kept increasing slowly during training, but still found it hard to decrease the validation loss. Does it mean I need a model with more capacity?

Answer 1

Loosely speaking, it means very likely that you data augmentation is useless . With other words, in your augmentation there is no new information for the neural network to learn. Increasing capacity of your deep net will results into more severe overfitting. I suggest

1. to overthink the way you augment your data (different strategy)
2. try to get hold of more data. Try to add dropout and L2-regularization to suppress overfitting.