Heuristics for parameters: patch size, halo, and rescale factor

[qin-yu]

I'm planning the following enhancements to PlantSeg that could significantly improve its usability:

• #251

  PlantSeg used a fixed halo that appeared optimal but was potentially influenced by a combination of incorrect padding implementation and a specific input image distribution for testing. Yet we might still be able to assume that input images remain friendly in the future, and therefore reduce the halo for practical reasons. I plan to enable this in ModelZoo because it is for all CNNs.

  Related discussion:

  • #248 (comment)

• Provide a maximum patch size based on GPU constraints and the decided halo

  Given the GPU resources and the halo dimensions determined, it would be beneficial to define a maximum patch size. I plan to prioritise size in z since we already figured out a theoretical and reliable but possibly oversized halo. e.g. for images with z < 128, I prefer to avoid padding in z dimension. As this requires information of input images as well, I plan to enable this in prediction modules.

  Related PR:

  • #307

• Provide a rescale factor based on voxel size

  I plan to make it a part of rescale widget.

  Related issues:

  • #189
  • #217

[lorenzocerrone]

Let me know when you want to tackle the rescale, I have some other issues with that widget.