Strong tiling artifacts

[Buglakova]

When I run the network prediction, the result has strong tiling artifacts. I use quite big halo, but it doesn't help. I encountered this before when using U-Net that wasn't trained for long enough or didn't see enough ground truth, which is kind of the case when applying a pretrained network to different data. As a workaround, would be nice to have an option for smooth transition between tiles, like in the original U-Net publication, where within the halo the weight of the tile in the result falls linearly towards the edge of the tile.

Here is an example:

Furthermore, it propagates to the superpixels (here GASP run on the wrong channel, but still you can see clearly the squares)

[qin-yu]

I'm looking into it and believe it relates to #205

[qin-yu]

Fixing #205 doesn't help, i.e. changing size of halo doesn't improve prediction, which indicates more serious problems. Thus #220 and wolny/pytorch-3dunet#113

Progress:

1. Halo implementation is fixed, which already greatly improved the prediction (Fix halo implementation and tiling artefact wolny/pytorch-3dunet#113 (comment))
2. Intensity normalisation doesn't cause problem at least for ovules (training data) and mouse (another dataset).
3. Batch norm fix hallucinations created by group norm in mouse embryo (I mean you need to train with batch norm again).

[qin-yu]

For PlantSeg:

• Halo implementation and intensity normalisation outside networks are fixed by Fix halo implementation and tiling artefact #220
• Batch norm with default track_running_stats = True should be used in models instead of group norm