How to train the global step and local step respectively? about cascadepsp

[hkchengrex]

Only one refinement module is trained. It is then used in the global/local step. They are not trained separately.

[GOZGOXGOYGO]

Thanks a lot, I get it. If I train the network on very high resolution images, could you please tell me what do I need to do to modified the onlinetransformdataset? Do I still need to use random crop 224x224?

[hkchengrex]

I haven't tried that but I suspect that you would need something more complicated.
At least you would need to use RandomResizedCrop instead of RandomCrop to get global context sometimes. That might lead to an overwhelming amount of useless crops (all foreground/background) and they should be discarded.
Moreover, when you do have high-res training data, one suggestion is to actually train the two steps separately -- one using downsampled large crops for the global step, and the other using small crops for the local step. This is not implemented in the current repo but might worth a try.

[GOZGOXGOYGO]

I see. I notice that the test dataset named "BIG_PSPNet_SS" you provided is large enough. Could I know how did you train the two steps on this dataset?

[hkchengrex]

We did not train the model on this dataset. It is too small and is for testing purposes only.

[GOZGOXGOYGO]

So, if I want to train on this dataset. I need to firstly generate the input data for global step. And use the output of global step to continuely generate input data for local step. Is that right?

[hkchengrex]

That's one way.
Another way is to generate the inputs for the global/local steps separately and independently by perturbing the ground-truth.

[GOZGOXGOYGO]

May I ask whether using the groundtruth-based perturbation results as "seg" or using the network-based prediction results as "seg" is better for training the two step? Have you compared these two situations?

[hkchengrex]

For our case (training two steps as a single module), it is better to use perturbation because we can generate much more training pairs with variations. I would suspect that it will also be better to use ground-truth perturbation even when you train them separately.