search

holoviz-topics / EarthSim

Tools for working with and visualizing environmental simulations.
https://earthsim.holoviz.org
BSD 3-Clause "New" or "Revised" License
65 stars 21 forks source link

Grabcuts minimum tolerance behavior #289

Closed kcpevey closed 5 years ago

kcpevey commented 5 years ago

If I run grabcuts once and get this:

image

If I just change the minimum tolerance from 0 to 500 and click Filter Contour, I get this:

image

This isn't intuitive to me - what is going on here? Is Filter contour running more iterations? Even if it where, this still wouldn't be intuitive.

kcpevey commented 5 years ago

I saw the above behavior a couple of times. I just recreated it with two new lines and just ran update contour once and now the results from very similar lines look like this:

image

I've been noticing erratic behavior like this from grabcuts quite a bit but I haven't been able to nail down the cause.

philippjfr commented 5 years ago

I've been noticing erratic behavior like this from grabcuts quite a bit but I haven't been able to nail down the cause.

Is this recently or has it always been acting odd for you?

kcpevey commented 5 years ago

I've always thought something was off. Unfortunately I reported a lot of it in a PR (https://github.com/pyviz/EarthSim/pull/186) instead of an issue. I'm sure some of it just the grabcut algorithm itself being finicky, but the issue above seems like it might be on our side.

Since we're on the topic, we need to switch the foreground and background specification. In a highly disjointed delta, grabcuts performs much much better when the the red marker is used to specify the water (because of the assumptions of the algorithm and which one is being used as 'source'. I haven't made an issue for this since I wanted to check to see if that was true for all images or if we just need to change the blue marker to a non-water color and explain that if you don't achieve good results, try and reverse them.

philippjfr commented 5 years ago

This helped me uncover an issue in the memoization of a DynamicMap. What should happen is that Update Contours should cache the result of the grabcut output and filter/simplify should only operate on the cached output of the algorithm. What was happening instead was that due to bad hashing behavior the cache wouldn't be used, resulting in the grabcut algorithm being rerun and I guess because it's unstable in some cases the output changes between runs, producing suboptimal results. I'll push a fix to holoviews and then you can check if the behavior has improved.

philippjfr commented 5 years ago

Here's the PR with a fix: https://github.com/pyviz/holoviews/pull/3490

kcpevey commented 5 years ago

This has improved.