Have you tried any type of stochastic pooling for regularization similar to dropout? about synthtext

[biggerlambda]

@cjnolet I think just the proposals without the postprocessing and tinkering with box dimensions leads to poor recall (see image below). @ankush-me can you confirm?

[cjnolet]

@biggerlambda, you are correct and I have trained both the filtering classifier and the bounding box regressor model. 25% is extremely low recall for the FCRN output which is making me thinking I'm doing something wrong, my network is overfitting, or I'm not training it for long enough.

[ankush-me]

From the gray curve above, multi-scale FCRN without any other post-processing gets 85% maximum recall on ICDAR. Precision can be improved through non-maximal suppression.

[cjnolet]

@ankush-me, I understand tyour answer to @biggerlambda's question/suggestion but my original question in this ticket still remains unanswered. Have you ever tried using stochastic pooling? Did you find that it did not help your convolutional layers?

Also, how many epochs did you need to train your SynthText dataset on?

[ankush-me]

I haven't tried any stochastic pooling.

I think it converged within 4-5 epochs.

(Closed in error -- feel free to comment if not resolved).