Why do MSER results have overlapping pixels

Question

First, I'm using opencv MSER via matlab via 2 different methods:

Through the matlab's detectMSERFeatures function. This seems to call the opencv MSER (via the call to ocvExtractMSER in the detectMSERFeatures function)
Through a more direct approach: opencv 3.0 wrapper/matlab bindings found in https://github.com/kyamagu/mexopencv

Either way, I can get back lists of lists of pixels (aka regions) that I imagine are a translation of the opencv MSER::detectRegions 2nd arg, "std::vector< std::vector< Point > > &msers"

The result can end up a list of multiple regions each region with its own set of points. However, the points of the regions are not mutually exclusive. In fact, they typically, for my data in which the foreground is typically roundish blobs, tend to all be part of the same single connected component. This is true even if the blob doesn't even have any holes (I might understand if the regions corresponded to contours and the blob had holes).

I'm assuming that this many-region-to-one mapping of regions to even a solid blob is due to opencv's MSER, in its native C++(?) implementation, doing the same but I confess I haven't verified that (but I surely don't understand it.)

So, does anybody know why MSER would yield multiple overlapping regions for a single solid connected component? Is there any sense to choosing one and if so how? (Right now I just combine them all)

EDIT - I tried an image with one blob which then I replicated to have a single image where the left half was the same as the right (each half being the same, each with the same blob). MSER returned 9 lists/regions all corresponding to the two blobs. So, I wold have to do connected component analysis just to figure out which subsets of the regions belonged to what blob and so apparently there can't be any straightforward way to choose a particular subset of the returned regions that would give the best representation of the two blobs (if such a thing was even sensible if you knew there was just one blob as per my last pre-edit question)

The picture below was made by plotting all 4 regions (lists of points) returned for my single blob image. The overlay was created by:

obj = cv.MSER('MinArea',20,'MaxArea',3000,'Delta',2.5);
[chains, bboxes] = obj.detectRegions(Region8b)
a=cellfun(@(x) cat(1,x{:}),chains,'UniformOutput',false) % get rid of extra layer of cells that detectRegions seems to give it.
% b=cat(1,a{:}); % all the regions points in a single list.  Not used here.
ptsstrs={'rx','wo','cd','k.'};
for k=1:4 
    plot(a{k}(:,1),a{k}(:,2),ptsstrs{k},'MarkerSize',15); 
end

So, you can see they overlap but there also seems to be an order to it where I think each subsequent region/list is a superset of the list before it.

If your blob has multiple gray levels, MSER might return multiple regions depending on how they are distributed. — dhanushka
What's the point if it doesn't differentiate between regions and the connected components they belong to? Maybe I have this wrong because i haven't used the actual opencv lib version and seen the raw MSER output, but if its like what I've seen, it returns a list of region points without any further info. — user3288226
Given one blob with multiple returned regions, what's the right one to use or is that a subjective thing. The results look like its returning what was computed along the way (going through the level sets or perhaps pruning) — user3288226

Tim Tim · Accepted Answer · 2018-03-22T16:49:01

"The MSER detector incrementally steps through the intensity range of the input image to detect stable regions. The ThresholdDelta parameter determines the number of increments the detector tests for stability. " This from Matlab help. It's reasonable that you find overlap and subsets. Apparently, the region changes as the algorithm moves up or down in intensity.

Why do MSER results have overlapping pixels

1 Answers