Question about num_iterations parameter in connected_components_with_gradients

BrainImageAnalysis / instance-loss

Official implementation of instance loss functions, a family of instance-level segmentation and detection loss functions, to improve semantic segmentation and detection of multiple instances in biomedical images

Apache License 2.0

1 stars 2 forks source link

Question about num_iterations parameter in connected_components_with_gradients #3

Closed aymuos15 closed 2 weeks ago

aymuos15 commented 2 weeks ago

Based on what has it been set to 75?

febrianrachmadi commented 2 weeks ago

Hi,

This is explained in the paper's Appendix B, written below.

If N is sufficiently large, then, after the iterations, each connected component (instance) will have a unique integer ID. Based on our experiments, a good trade-off between the number of iterations (N) and the computation complexity is setting N equals to half the image’s maximum resolution. For example, for an image with 512 × 128 resolution, it would be 512 / 2 = 256.

Thus, 75 is just a default number for N.

aymuos15 commented 2 weeks ago

Thank you very much!