search

system123 / SOMatch

A Framework for Deep Learning-based Sparse SAR-Optical Image Matching
MIT License
48 stars 17 forks source link

how to train correlation model with sen12 #6

Closed yaboliudotug closed 3 years ago

yaboliudotug commented 3 years ago

Thanks for sharing your work. I follow your code and train a correlation model with the sen12 dataset. However the loss I trained is not converged. Are there any training skills or key hyperparameters? We also looking forward to your pretrained model. Thanks.

system123 commented 3 years ago

I just provided the SEN12 dataloader as an example, I have not spent much time testing this model on medium resolution imagery so I am not sure about the hyperparameters. You will most likely need to perform a hyperparameter optimization to find them.

The loss decreases slightly but you'll need to smooth it to see, a better measure is the quality (peakiness) of the correlation heatmaps. nThe optimization problem in this case is quite difficult to obtain a strongly decreasing loss as the output is a single point with all other areas being zero – mse, and even L1 losses are not ideal for such a problem but they do work. I trained the models based on a fixed number of epochs instead, this number was determined through experimentation and evaluating the quality of the training heatmaps.