Warpfield as a "pointcloud registration" tool?

[danielhannuschke]

Hello,

first of all I wanted to say how amazing warpfield is! I've been using it to register some multiplex IF and it works amazingly well. Thanks!
I was wondering if this approach could also be used to perform a sort of pointcloud registration approach. Essentially what I want to do at some point is to warp a dense source pointcloud to a target volume. I realize this is more of a surface deformation problem but since I also care about the displacement of the points inside my object I was thinking that the warpmaps from this tool could be a promising road.
For context I am working with zebrafish embryos so for the point cloud, think of a half-sphere filled with cells, and the deformation happening is an invagination at the center of the base circle. I'm happy to share some example images, e.g. I tried just applying warpfield to the embryo segmentation mask, but obviously it needs some intensity gradients to work properly.

I also had a minor thing: Is it also possible to save and load the WarpField objects at the moment? I couldn't find anything in the code...

Any ideas or pointers would be appreciated!

[bjudkewitz]

Hi @danielhannuschke , great to hear that warpfield has been working for you! If you like to, you are welcome to share some registration videos in the show-and-tell channel. We are genuinely curious about various – often unforeseen – use cases.

Pointcloud... we can apply any warp field to points, but I don't think warpfield will ever support registering point clouds directly. Not because I think there is no need for it, but the underlying approach would be so different that it would merit a separate tool.

There are of course workarounds, if you find a way to convert your coordinates to dense volumes. You could, for example, create a 3D histogram of your points and smooth them, as long as the outcome looks somewhat similar to the reference volume (which you could prepare in analogous ways).

I also had a minor thing: Is it also possible to save and load the WarpField objects at the moment? I couldn't find anything in the code...

There is no convenience function for it yet, but you could, for example, write sth like this:

import cupy as cp
import numpy as np
from warpfield.register import WarpMap

def wm_to_dict(wm):
    to_host = lambda x: cp.asnumpy(x) if isinstance(x, cp.ndarray) else x
    keys = ("warp_field", "block_size", "block_stride", "ref_shape", "mov_shape")
    return {k: to_host(getattr(wm, k)) for k in keys}

def dict_to_wm(d):
    return WarpMap(d["warp_field"], d["block_size"], d["block_stride"],
                   tuple(d["ref_shape"]), tuple(d["mov_shape"]))

(and save/load the dict with your favorite tool, e.g. daio.h5)

See answer below

[bjudkewitz]

Since it was a simple change, I added the .to_h5 method and .from_h5 classmethod to WarpMap to save/load instances to/from HDF5 files. Works in warpfield 0.2.1+. See WarpMap documentation for more details.

Example:

wm.to_h5('test.h5')
# some time later:
wm2 = warpfield.register.WarpMap.from_h5('test.h5')

[bjudkewitz]

Since you didn't answer I'll take the optimistic interpretation that it worked for you.