Why is there a val and train folder with images?

[heblol]

Dear Guru and Nadine,

Great work reproducing the original work. Currently I am using your code for a study, but I have a question about training.

I have a question why the train.py uses both images in the /val and /train folder? I seems the /val has not a use when the latent_images is turned off? Also both folders don't have the same amount of images, is there a reason for this? I want to train the W-Net on the Kaggle Wheat Detection dataset, should I copy the training data in both the folders?

Thank you for sharing the code,

Thank you in advance,

Sincerely,

Ivo Chen

[Guru-Deep-Singh]

Hi Ivo,
No the only use for the val images is to visualize the improvement of the model after every epoch. In the original paper author uses only near to 11k training images while Pascal VOC 2012 dataset has near to 17k images. You can just put few images in the validation folder while put the complete dataset in train.
Good Luck
Guru

[heblol]

Thank you for replying.

I have another question. In your Medium post about this repository, you don't mention the post processing steps.

After investigating the code it seems you implemented the CRF smoothing step, but I cannot find the Hierarchical Segmentation step.

Is it an correct assumption your have implemented CFR Smoothing correctly, but have not implemented hierachical segmentation? Is there a special reason why you did not address this in the post on Medium?