Instance Segmentation using pix2pix

[seekingdeep]

@lmxyy @junyanz Hi there,

i am trying to use pix2pix to train an instance segmentation model, My input images that are used in training are 256*256 crops of my original 1024*1024 images.

The issue is that:

• if using 256*256 crops in training, the model is not converging and the fid is somewhat bad. Because the 256*256 crops are too small, and the information of the "big picture" is somewhat lost.

• When trying to using the original 1024*1024, it requires high memory when running get_real_stat.py.

Possible solution: # 44 is to store the image tensor on the disk and then modify the dataloader to restore the tensors from the disk.

i think using the HD original images 1024*1024 in training should get better results, Waiting for your reply

iteration 185000, Prediction:

Groundtruth:

[lmxyy]

I think you could try the original resolution. Storing the image tensor on the disk and then modify the data loader to restore the tensors from the disk is a good solution when running the get_real_stat.py is a good solution and deserves a trial.