fix subidx to compute motion correction template from a movie slice #1
Conversation
| num_splits=num_splits_to_process, shifts_opencv=shifts_opencv, nonneg_movie=nonneg_movie, gSig_filt=gSig_filt, | ||
| use_cuda=use_cuda, border_nan=border_nan, var_name_hdf5=var_name_hdf5, is3D=is3D, | ||
| indices=indices) | ||
| indices=indices, subidx=None) |
There was a problem hiding this comment.
FYI subidx=None by default in motion_correction_piecewise, but no harm in stating that explicitly here
There was a problem hiding this comment.
Right, and I don't know why they had subidx=subidx in the function call, as it tells motion correction to run just on the specified slice (that's what was making subidx dual function, one the one hand to constrict frames from which to compute the template, and on the other hand by computing motion correction on this only). Writing this I think they maybe never had the use case of wanting to apply a template on a superset of frames, if it makes sense! So maybe they just put that here to say "if I have a movie and just want to do MC from frame A to frame B, I can do so" and it impacted both template computation and MC itself.
| @@ -1,7 +1,7 @@ | |||
| channels: | |||
| - conda-forge | |||
| dependencies: | |||
| - python >=3.10 | |||
| - python =3.9 | |||
There was a problem hiding this comment.
great! we should add specific versions for all dependencies as we make improvements (e.g. I know some older versions of numpy don't work)
When motion correcting a series of recordings (i.e., concatenated movies), in some cases there is a shift between movies' FOVs, and the default template/reference frame computed by averaging 50 frames uniformly sampled across the series is spoiled. Concretely, it's like averaging (for two movies of equal duration) half frames with a "resting position" (x, y) and half frames with a shifted resting position (x + a, y + b), resulting in a blurry / dual state template and eventually to failed motion correction.
One solution is to compute the template from one movie only, and that's what the argument "subidx" is meant to do: define a movie slice from which to sample the 50 frames. Only, 1) this argument wasn't taken into account, and 2) it also affected the output movie (only output the chosen slice), making it useless.
This PR fixes "subidx", and allows for a successful template/reference frame computation (using the exact same sampling and spatial filtering steps as built-in) from one movie only, critically enabling the successful use of NoRMCorre on concatenated movies.