Input data process

[GeoSur]

Hello Mr. Thomas, I have a few questions during the data processing:

• Is the dataloader operation based on potentials/probabilities? First, initialize the potentials/probability of each point and then iterate through all the input data as it is updated. So how should we ensure that all the training data is fed into the network?
• Why do we need a calibration function in the Sampler class? it seems that the changes of in_radius and first_subsampling_dl will influence the calibration function. I can't understand these operations clearly, if I want to know the data feeding process and the changes of these data in the network, what should I learn about it?
• When a dataset has 3 validation areas, the mIou of the validation process is sometimes not displayed, is it due to a too short evaluation process? As the photo shows...

[GeoSur]

Hello! @HuguesTHOMAS

• I have been troubled by these problems recently and hope to your guidance. I found that the original_ply data will be grid_subsampled twice before being fed into the network, the first sampledl is what we config this process we got input_trees, and the second is in_radius/10this process we got pot_trees, why should we do this operation? We update the potentials from pot_trees, and get the input data from input_trees. Is this for faster data loading?
• And about the load_subsampled_clouds function: both of these note that 'Only necessary for validation and test sets', but I think only Reprojection indices doing what it noted.