ZJULearning / pixel_link

Implementation of our paper 'PixelLink: Detecting Scene Text via Instance Segmentation' in AAAI2018
MIT License
768 stars 255 forks source link

Validation Method #123

Open bajarno opened 5 years ago

bajarno commented 5 years ago

What validation method is used to select the model for which results are reported on the ICDAR2015 test set?

I was unable to find how a model is selected while reading the paper and looking through the source code. It seems like the complete training set is used for training, so no data is left for a separate validation set.

I am trying to reproduce the results using PyTorch. My model is performing reasonably well, but the results are rather unstable over different epochs. Some validation method is needed to properly select a good model and experiment with hyperparameters while preventing overfitting on the test set.

Thank you!