Add widget: remove objects using a probability map

[qin-yu]

A generic pmap-guided object removal widget

Instead of adding an option to each "boundary -> segmentation" algorithm as discussed in #174, I made a separate widget for Napari GUI pipeline for generic usage. This independent design avoids bundled computation within other functions such as GASP, allows easy trial-and-error strategy, accepts flexible input pmaps/images and threshold, and is useful in unexpected use cases.

I came up with three designs and decided to implement only one. I now add this to Napari, and will think about the best way to do it in the other UI.

1. Method 1: Threshold foreground to a binary mask and multiply with the segmentation pixel-wise ❌
   1. Pros: very fast; false positive parts will be deleted within objects if foreground quality is high
   2. Cons: depending on the quality of foreground segmentation, details may be lost
2. Method 2: Delete instances based on the centroids' value in foreground ❌
   1. Pros: more educated guesses than 1, preserves details
   2. Cons: merged instances or merged false-and-true instances will be deleted together; not helpful if instances are not star-convex, e.g. C shaped false positives around objects will persist
3. Method 3: Delete instances that has, say, 90% overlap with the background (I believe it's the best way ✔)
   1. Pros: considering the foreground probability of each pixel in an instance, the algorithm is well-informed
   2. Cons: slightly slower than 1 & 2 but after careful design and simple tests I believe it handles segmentation/superpixels at a scale of 100k instance in minutes

Example Usage

As a standalone widget it not only helps reduce unnecessary computation but also can be flexibly used between different steps. For example, the following image shows how an instance connected to a false positive would be deleted if we simple use GASP followed by method 3:

To have a perfect segmentation, users can apply a method 3 foreground filter before and after GASP like this:

Starting the a watershed output:

Apply the new filter:

Run GASP:

Apply my new filter again:

[qin-yu]

It does work

[qin-yu]

I think I'm done with the addition of widget, works well:

[qin-yu]

A problem of only have foreground filter as a post processing:

[qin-yu]

Screenshots

Raw:

Foreground prediction:

Boundary P\prediction:

Watershed superpixels shown on boundary:

Watershed superpixels shown on raw:

Filtered superpixels on foreground:

Filtered superpixels on raw:

GASP ran on filtered superpixels:

Filtered GASP:

[qin-yu]

Hey @lorenzocerrone and @wolny

I added a widget that uses probability maps to filter objects in another image. It is very flexible but only available in Napari now. I believe the other less interactive interfaces don't need this level of flexibility, and I'll think about how to design the filter into the them later. Please have a look at this PR.

[wolny]

this looks great, nice job @qin-yu