Return nearest contour point + memory leak fix

[Ali2500]

This is related to issue #29 . Two key changes have been made:

1. A new return_nearest_contour option has been added to the distance, travel_time and extension_velocities functions (default to False). If true, the function returns an additional array which contains the coordinates of the nearest phi = 0 point for each point in the input array. For e.g., for a 2D image of size HxW, the returned nearest contour array will have size HxWx2.

2. The current code can leak memory because of the way in which Py_BuildValue is used. Specifically, using O in the format string causes the reference counter of the objects to be increased, so the returned array(s) are never freed. The solution is to use N instead, which doesn't increase the reference counter.

I didn't name the option 'image_gradient' since it returns the absolute coordinates of the nearest phi = 0 point, not the coordinate offset to it, but that can be changed easily.

I also wrote a simple interactive script that can be used to test the functionality. The code for it is pasted here (requires OpenCV, NumPy). It basically converts the generated distance array into an easily visualizable color map. Clicking anywhere on the image causes the nearest contour point to be marked with a green circle (the exact coordinates are also printed to console).

https://www.pastiebin.com/5d1a070f101a2

A simple test image on which it can be applied (change the file path in the script accordingly). It appears all black but there are some points with intensity 1:

[mrmotallebi]

@Ali2500 I don't know if you still remember your code but I tried to merge your changes to the head of this repo (mrmotallebi#1). However I am getting all 0s in some cases for the nearest_contour output (especially when part of the input array is masked out). Would you have any idea why?

[jkfurtney]

Can you give an example of what you are seeing and what the expected behavior is? I can revisit this branch and help with the merge. Out of curiosity, what is your use case? Thanks.

[mrmotallebi]

Can you give an example of what you are seeing and what the expected behavior is? I can revisit this branch and help with the merge. Out of curiosity, what is your use case? Thanks.

@jkfurtney Try this:

X, Y = np.meshgrid(np.linspace(0,15,15), np.linspace(0,15,15))
phi = np.ones_like(X)
mask = np.zeros_like(phi).astype(bool)
mask[0,:] = True
mask[:,0] = True
phi = np.ma.MaskedArray(phi, mask)
phi[10,10] = -1
dist, contour = skfmm.distance(phi, return_nearest_contour=True)

The distance is correctly calculated, but controur is all zeros.
(It was based on the fourth example in the docs .)