On-the-fly conversion of instances to semantic masks/affinities

[Mohinta2892]

Hi Dani,

For an extremely large dataset, the converting all instance labels to masks can be tedious and lead to data redundancy.

It would be easier with a method (as an augmentation) in the pipeline, which when called can generate masks when reading in labels from the input dataset on the fly during training.

If it can be implemented, that would be amazing. Thanks!

Best,
Samia

[danifranco]

Hi Samia,

There is always and advantage and disavantage on the choosen method. Is true that you need disk when the masks are created, but the advantage is that you don't waste time in creating the masks on the fly, which for long training settings it reduces significatively the training time. Depending on the complexity of the masks to be created the procedure is not simple and may take too much time to generate them on the fly.

I mean, I see your point but there is always pros and cros in all settings... We can definitely do it, so we can leave this issue in case any of us can take a look to it. Also, feel free to make a PR if you want!