Mailman 3 March 2012 - scikit-image

Re: Normalised cross correlation
by Tony Yu March 30, 2012

March 30, 2012

On Fri, Mar 30, 2012 at 9:55 AM, Tony Yu <tsyu80(a)gmail.com> wrote: > > > 2012/3/27 Stéfan van der Walt <stefan(a)sun.ac.za> > >> On Mon, Mar 26, 2012 at 6:36 PM, Tony Yu <tsyu80(a)gmail.com> wrote: >> > Thanks for checking. What do you mean by "correct position in the >> > correlation"? >> >> It looked like you took the result of np.correlate, and then >> normalized those coefficients. I wasn't sure (didn't verify) whether >> the normalization was computed for the correct position in the >> np.correlate result. >> > > Yes, you're correct---not sure why I was confused by before. > > >> > Hmm, I was under the impression that half-template matches would be >> > appropriately penalized. >> >> That explanation seems reasonable--such behaviour seems totally sane. >> >> Stéfan >> > > > I added some commits to the template PR<https://github.com/scikits-image/scikits-image/pull/100>. > I changed the interface: `pad_output` is replaced by `pad_input`, which > defaults to False. When True, the input image is padded with a border that > is half the template dimensions on each side, and the padding value is the > image mean, as Mike suggested. Note that this function still does not > produce results similar to Matlab's; no one made the case for it, so I left > it out (and as Stefan pointed out: the user can manually pad the input if > such a result is desired). Are people happy with this interface? > > I added a simpler example plot<https://github.com/scikits-image/scikits-image/pull/100/files#diff-0>to the gallery, but there should probably only be a single example. > Personally, I prefer the simpler plot. (The more complex plot<https://github.com/scikits-image/scikits-image/pull/100/files#diff-1>, > however, is more "realistic" since it has noise and uses a peak detector.) > Any objections to removing the complex plot? > > -Tony > Actually, I decided I like this third example<https://github.com/scikits-image/scikits-image/pull/100/files#diff-2>better. (I'm just trying to see how many commits I can get into this PR before it's closed.) -Tony

1 0

Re: Normalised cross correlation
by Tony Yu March 30, 2012

March 30, 2012

2012/3/27 Stéfan van der Walt <stefan(a)sun.ac.za> > On Mon, Mar 26, 2012 at 6:36 PM, Tony Yu <tsyu80(a)gmail.com> wrote: > > Thanks for checking. What do you mean by "correct position in the > > correlation"? > > It looked like you took the result of np.correlate, and then > normalized those coefficients. I wasn't sure (didn't verify) whether > the normalization was computed for the correct position in the > np.correlate result. > Yes, you're correct---not sure why I was confused by before. > > Hmm, I was under the impression that half-template matches would be > > appropriately penalized. > > That explanation seems reasonable--such behaviour seems totally sane. > > Stéfan > I added some commits to the template PR<https://github.com/scikits-image/scikits-image/pull/100>. I changed the interface: `pad_output` is replaced by `pad_input`, which defaults to False. When True, the input image is padded with a border that is half the template dimensions on each side, and the padding value is the image mean, as Mike suggested. Note that this function still does not produce results similar to Matlab's; no one made the case for it, so I left it out (and as Stefan pointed out: the user can manually pad the input if such a result is desired). Are people happy with this interface? I added a simpler example plot<https://github.com/scikits-image/scikits-image/pull/100/files#diff-0>to the gallery, but there should probably only be a single example. Personally, I prefer the simpler plot. (The more complex plot<https://github.com/scikits-image/scikits-image/pull/100/files#diff-1>, however, is more "realistic" since it has noise and uses a peak detector.) Any objections to removing the complex plot? -Tony

1 0

Re: [OT] pass a C callback function to a cythom module via python
by Nadav Horesh March 28, 2012

March 28, 2012

Thank you. I looked at the example and I need some time to digest it. I'll take a loser look over the weekend. Nadav. בתאריך 28 במרץ 2012 00:03, מאת Stéfan van der Walt <stefan(a)sun.ac.za>: > Hi Nadav > > I just had a chat with Dag, a core Cython contributor. He mentioned > to me another alternative, which is to encapsulate the callback > function in a class: > > > https://github.com/dagss/cython-pydata12/blob/master/examples/oointegrate/i… > > You can import this class from Python, and pass it along as the > callback object. There may be some work on improving this > functionality in the future, but for now I think that's the cleanest > route (and helps us to avoid ctypes!). > > Stéfan > > > On Tue, Mar 20, 2012 at 1:02 AM, Nadav Horesh <nadavh.horesh(a)gmail.com> > wrote: > > > > Can you explain in more details how can I implement it? Any case I am > > attaching an scheme that may clarify my intend. Soon I'll post a code > > demonstrating the problem. > > > > Thank you, > > Nadav, >

1 0

Re: Platform for filters development
by Nadav Horesh March 28, 2012

March 28, 2012

Thank you. I am over-busy these days so I'll look at it ot the week-end. Nadav. בתאריך 27 במרץ 2012 06:08, מאת Tony Yu <tsyu80(a)gmail.com>: > > > On Wed, Mar 21, 2012 at 12:05 PM, Nadav Horesh <nadavh.horesh(a)gmail.com>wrote: > >> (I think I missed the attachment) >> >> >> Attached here a concept code showing my attempt to produce in improved >> (over ndimage.generic_filter) platform for filters development. The code is >> very preliminary and needs a lot of extension as well as rewrite. My main >> questions are: >> >> 1. Is there is already a mature package for this purpose? >> 2. Is there is need for such a module inside scikits? >> >> Any comment is welcome. >> >> Nadav >> > > Hey Nadav, > > I had some problems building this example. (Actually, I couldn't even > unzip it at first.) In case anyone else wants to try this out, I posted a version > of your package <https://github.com/tonysyu/filter_iter> to my github > account with some build modifications. I think the build requires Cython > 0.15 (or at least, I think so: I was getting some build errors with 0.14 > and upgrading seemed to fix it (although I may have made other changes > along the way).) > > I haven't yet had time to play around with the code, so I don't have > anything constructive to say about the code yet. > > Cheers, > -Tony >

1 0

Re: [OT] pass a C callback function to a cythom module via python
by Stéfan van der Walt March 27, 2012

March 27, 2012

Hi Nadav I just had a chat with Dag, a core Cython contributor. He mentioned to me another alternative, which is to encapsulate the callback function in a class: https://github.com/dagss/cython-pydata12/blob/master/examples/oointegrate/i… You can import this class from Python, and pass it along as the callback object. There may be some work on improving this functionality in the future, but for now I think that's the cleanest route (and helps us to avoid ctypes!). Stéfan On Tue, Mar 20, 2012 at 1:02 AM, Nadav Horesh <nadavh.horesh(a)gmail.com> wrote: > > Can you explain in more details how can I implement it? Any case I am > attaching an scheme that may clarify my intend. Soon I'll post a code > demonstrating the problem. > > Thank you, > Nadav,

1 0

Re: Normalised cross correlation
by Stéfan van der Walt March 27, 2012

March 27, 2012

On Mon, Mar 26, 2012 at 6:36 PM, Tony Yu <tsyu80(a)gmail.com> wrote: > Thanks for checking. What do you mean by "correct position in the > correlation"? It looked like you took the result of np.correlate, and then normalized those coefficients. I wasn't sure (didn't verify) whether the normalization was computed for the correct position in the np.correlate result. > Hmm, I was under the impression that half-template matches would be > appropriately penalized. That explanation seems reasonable--such behaviour seems totally sane. Stéfan

1 0

Re: Platform for filters development
by Tony Yu March 27, 2012

March 27, 2012

On Wed, Mar 21, 2012 at 12:05 PM, Nadav Horesh <nadavh.horesh(a)gmail.com>wrote: > (I think I missed the attachment) > > > Attached here a concept code showing my attempt to produce in improved > (over ndimage.generic_filter) platform for filters development. The code is > very preliminary and needs a lot of extension as well as rewrite. My main > questions are: > > 1. Is there is already a mature package for this purpose? > 2. Is there is need for such a module inside scikits? > > Any comment is welcome. > > Nadav > Hey Nadav, I had some problems building this example. (Actually, I couldn't even unzip it at first.) In case anyone else wants to try this out, I posted a version of your package <https://github.com/tonysyu/filter_iter> to my github account with some build modifications. I think the build requires Cython 0.15 (or at least, I think so: I was getting some build errors with 0.14 and upgrading seemed to fix it (although I may have made other changes along the way).) I haven't yet had time to play around with the code, so I don't have anything constructive to say about the code yet. Cheers, -Tony

1 0

Re: Normalised cross correlation
by Tony Yu March 27, 2012

March 27, 2012

2012/3/26 Stéfan van der Walt <stefan(a)sun.ac.za> > Hi Tony > > On Sat, Mar 24, 2012 at 4:03 PM, Tony Yu <tsyu80(a)gmail.com> wrote: > > zero, but round-off errors are causing it to go negative. If that's the > > case, then it's an easy fix to check that the argument is positive, but > > someone should check my math to make sure the the code matches the > equation. > > I checked the paper vs your code, and the code looks correct. The > only thing I didn't check was whether the normalization was done on > the correct position in the correlation. > Thanks for checking. What do you mean by "correct position in the correlation"? > > > Padding *the input* sounds like a good idea, but I think it should be > done > > in such a way that the output is (M, N)---i.e. same size as the input > image. > > I'm not a huge fan of the Matlab output size (but maybe some people > expect > > this output?). > > If the user wants to pad the input, they should probably do that > themselves beforehand. Half-template matches feel sketchy (and where > do you stop: when at least half the template matches?). Also, with > the new proposed padding in numpy, this should become a lot easier. Hmm, I was under the impression that half-template matches would be appropriately penalized. Say we have a template of a face: If it perfectly matches a patch of an image, the result will have a match value of 1 at that location. Now, if half the face matches perfectly but the rest is garbage, then the result might get a value of 0.5 (I'm not sure what the value would actually be). To me, it doesn't really matter whether the face is half-obstructed by a shadow (but still fully in the image), or if it's half cropped by the image edge; they should both have values around 0.5. Unfortunately, the exact score would depend on what value was used to pad the image (Mike's suggestion of the average seems preferable, but nothing's going to be perfect). I agree that it'd be better for users to take care of the padding themselves, but I think it's a common-enough use-case that we should make it easy to pad with the image average (although turn it off by default). -Tony > > Stéfan >

1 0

Re: Normalised cross correlation
by Tony Yu March 27, 2012

March 27, 2012

2012/3/26 Stéfan van der Walt <stefan(a)sun.ac.za> > On Sat, Mar 24, 2012 at 5:09 PM, Tony Yu <tsyu80(a)gmail.com> wrote: > > Just to follow up: This would be my preferred way to pad the image. Here, > > the maximum in the output corresponds to the center of the best > > template-match. > > Just as a point of interest, there should soon be some padding > capabilities in numpy: > > https://github.com/numpy/numpy/pull/198 > > Stéfan > This looks like a great new feature. Thanks for the heads-up. -Tony

1 0

Re: Normalised cross correlation
by Stéfan van der Walt March 26, 2012

March 26, 2012

Hi Tony On Sat, Mar 24, 2012 at 4:03 PM, Tony Yu <tsyu80(a)gmail.com> wrote: > zero, but round-off errors are causing it to go negative. If that's the > case, then it's an easy fix to check that the argument is positive, but > someone should check my math to make sure the the code matches the equation. I checked the paper vs your code, and the code looks correct. The only thing I didn't check was whether the normalization was done on the correct position in the correlation. > Padding *the input* sounds like a good idea, but I think it should be done > in such a way that the output is (M, N)---i.e. same size as the input image. > I'm not a huge fan of the Matlab output size (but maybe some people expect > this output?). If the user wants to pad the input, they should probably do that themselves beforehand. Half-template matches feel sketchy (and where do you stop: when at least half the template matches?). Also, with the new proposed padding in numpy, this should become a lot easier. Stéfan

1 0