Roman Grothausmann

Yes, I think, considering your findings, that the initial gradient computation leads to a very large gradient (possibly even in false direction) "due to termination of the fast marching" just...

> I've managed to reproduce the situation where the initial gradient is zero. It happens when the values in the speed image are very low, leading to neighbours remaining unevaluated...

> I've implemented a safe interpolator and some tests. The code is in the SafeInterpolator branch of my repo. Many thanks @richardbeare, I wanted to try and had a look...

Tested with ITK @ https://github.com/romangrothausmann/ITK/commit/47768d88e02038cd89c9dc89208a5299b0cb5d65 using https://github.com/richardbeare/ITKMinimalPathExtraction/commit/2f54421df763e8121c5a3f828e630c85769c81c0 for `ITKMinimalPathExtraction`. Setting the special interpolator works (https://github.com/romangrothausmann/ITK-CLIs/commit/b6b9667392016c2f02797a1b65866b71201807aa). With that the gradient decent optimizer did start for our test case and got as...

Tested this with the same CLI and get the same results as with the implementation provided by @richardbeare for #61: https://gitlab.com/romangrothausmann/ITK-CLIs/pipelines/130368269

> It looks like some test baselines need to be updated and other tests are not generating baselines. > > @romangrothausmann do all tests pass locally? Sorry, only tested with...