[titusgriebel]

@anwai98

Motivation: In a scenario where a user wants to locally annotate a series of images using object classification, object features are not cached right now between sessions. This means that if the annotation process is split into several portions, significant waiting times occur, especially in data with larger object counts.

'skimage.transform.resize' took 94% of the object feature computation time: about 0.5 s per tile, so 8 s for a 2048 px field with the default tiling. 'torch.nn.functional.interpolate' takes about 0.01 s per tile. The values match, except in the 2 px edge of the resized embedding, where torch repeats the edge pixel and skimage mirrors it.

Thus, even without feature caching, recomputing features would become much less of a burden.

[anwai98]

Thanks @titusgriebel. I'll check this later today too (could also port this to bioimage_cpp's resampling function -- will review later anyways)

[constantinpape]

Makes sense from my end. @anwai98 feel free to merge later. We can use torch; as long as the results are numerically the same (2 px edge artifacts don't matter) . We don't need to check for alternatives.

[anwai98]

Thanks for the PR @titusgriebel. Looks good to me and CP. I"ll go ahead and merge this once the CIs pass!