search

scikit-image / skimage-demos

Note: this repo is not actively maintained, and may be out of date.
Other
37 stars 23 forks source link

ENH: New, expanded panorama demo w/MCP stitching! #4

Closed JDWarner closed 9 years ago

JDWarner commented 9 years ago

Take a look, @stefanv - the minimum cost path method to stitch images together works very well.

No more need for Enblend!

I limited the long axis on these pictures (which I have listed as CC-BY 4.0 right now, but could be persuaded to move to CC0) to 720 pixels. They're under 2MB in total, and ORB features take about 6 seconds total to extract from either set on my machine.

JDWarner commented 9 years ago

@stefanv if this looks good to you, I think it's ready to merge on my end.

Allowing "sliding" borders for MCP was a capital idea!

stefanv commented 9 years ago

This looks great, Josh! I think it would be good to add a way to export the final image to jpeg.

JDWarner commented 9 years ago

Done!

JDWarner commented 9 years ago

Don't merge quite yet, I'm going to simplify the MCP path seeds because the sliding border work so well.

JDWarner commented 9 years ago

Ruh roh, I found another challenge - to fix it requires tweaking the MCP cost function. Think I've got a fix.

JDWarner commented 9 years ago

@stefanv I think my work is done here!

The new cost array building function is necessary. I couldn't find any way to simplify it further. The results are now excellent and robust.

Basically, because the area surrounding the diff intersection was penalized so strongly, if the intersection border wasn't parallel with the top/bottom borders, the minimum-cost paths made a beeline to the corner nearest the border. This became evident when I allowed them to slide freely across the entire side. The new function gives slightly longer paths appropriate "breaks" in cost to be equally attractive.

stefanv commented 9 years ago

Remind me again why the area outside the diff is penalized?

JDWarner commented 9 years ago

If not, the MCP finds the trivial solution by skirting the entire diff region.

stefanv commented 9 years ago

I guess the question should be whether the correct constant background cost could save us all the effort.

JDWarner commented 9 years ago

No constant background will work this well; in that case you always have a contest between distance from the edge vs. quality of overlap. The MCP will choose essentially the midpoint between these, often resulting in visible exposure differences from vignetting. Trying to minimize this involves lowering the constant background toward the limit of zero... but you'll never get there; at some point it will snap to the trivial solution.

The only way to make it robustly choose the best path is to eliminate distance from the edge as a factor.

stefanv commented 9 years ago

@JDWarner How about if we use a distance transform as computed by scipy.ndimage.distance_transform_edt?

JDWarner commented 9 years ago

@stefanv I tried the various distance transforms. The distance function alone cannot be used, as it makes the incentive to go for the nearest corner even stronger.

If you forge ahead, they all suffer from two big problems:

It ends up being harder and more complicated than doing it the right way.

@JDWarner https://github.com/JDWarner How about if we use a distance transform as computed by scipy.ndimage.distance_transform_edt?

— Reply to this email directly or view it on GitHub https://github.com/scikit-image/skimage-demos/pull/4#issuecomment-104061650 .

stefanv commented 9 years ago

I am worried that this part of the tutorial will go right over people's heads. Let's give it one more "thinking try", and if we cannot come up with any ideas over the next day or to, merge it as-is.

JDWarner commented 9 years ago

Updated the notebook. I simplified and improved the readability (or split) several dense cells. One of the sections in the costs function was extraneous, and now it's removed.

There is no better way to do this I know of. It does look simpler now. This (95452f4) is the version included in skimage-tutorials/pulls/17.

stefanv commented 9 years ago

Great work--since I can't come up with a simpler way, let's use this one. And you get to explain it to the attendees :)