Mailman 3 September 2013 - scikit-image

binary_fill_holes, but creating separate regions inside each other
by Robin Wilson Sept. 21, 2013

Sept. 21, 2013

Hi, I am currently using various skimage functions to perform a customised edge-based image segmentation. Once I've got the outlines of my segments, I am using binary_fill_holes to fill them so I get a binary 'blobs' image, which I can then label, and use for processing. However, when I have an image like this 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 1 1 0 1 0 0 0 0 0 0 1 0 1 0 1 1 1 1 0 1 0 1 0 1 0 0 1 0 1 0 1 0 1 0 0 1 0 1 0 1 0 1 0 0 1 0 1 0 1 0 1 1 1 1 0 1 0 1 0 0 0 0 0 0 1 0 1 1 1 1 1 1 1 1 binary_fill_holes simply fills the entire, outer shape - and thus the shape in the middle gets lost. Is there any simple way to do a fill so that I get two regions: the outer ('donut-shaped') region and the inner region? Or will I have to develop and implement my own algorithm to do this? Any advice would be great, Thanks, Robin

2 1

Re: ndarray subclasses
by Josh Warner Sept. 21, 2013

Sept. 21, 2013

Indeed, marching cubes has anisotropic support presently. `random_walker` also does, though I think my original API for it is poor (only supports anisotropic `depth`, rather than true arbitrary anisotropy via `sampling`). I have an improvement/generalization PR for that in the works. I believe the most recent SLIC improvements also added a `sampling` parameter. On Friday, September 20, 2013 6:55:58 AM UTC-5, Almar Klein wrote: > > On Thu, Sep 19, 2013 at 12:03 PM, Almar Klein <almar...(a)gmail.com<javascript:>> >> wrote: >> > I suppose that the most important bit is that functions that support >> > anisotropy should look whether a "sampling" attribute is present in the >> > given array. >> >> I don't think we currently have any of these, but for now we can >> probably include a `sampling` argument to functions that support it. >> > > Mmm, you seem to be right, but you're *going* to :) The new marching > cubes algorithm has a sampling argument, and I think Josh spoke about > adding it so some morphological operators. I hope to do a PR on the MCP > algorithm soon, which will add support for anisotropy as well. > > > > I do not necessarily mean that a PointSet represents an image (although >> it >> > could), but more generally to for instance store the locations of >> detected >> > feature points. >> >> How do you think this would fit into the scope of image processing? >> (Asked out of curiosity, not at all to put the idea down.) >> > > That's a good point. I think a PointSet class fits image processing > because many image processing algorithms either accept or produce some form > of locations or vectors. I would still call it "image processing" when you > process the resulting locations/vectors. However, such algorithms probably > fall out the scope of scikit-image. > > So probably a better place for a PointSet class would be Scipy, but I have > a feeling they would not be interested in including it. > > > I am not related to the development of scikit-image but I guess its goal >> is to work on images in general and not get too specialised, for instance >> in Medical images. >> > > I agree with that. I think that the only parameter of interest to scikit > image is the "sampling" to deal with anisotropic arrays. So in that sense, > it should be sufficient that sciki-image provides an Image class to which > arbitrary attributes can be attached. Functions in scikit-image can the > check if the array has a "sampling" attribute, and use it. If it returns an > array with the same shape, it would be nice to also set the sampling on > that. > > > May I ask: if you were to add sampling information and spatial >> coordinates, such as an origin for the top-left pixel, how would you input >> that information into scikit-image, could you read it directly from the >> input files or would you need some user input? >> > > I am not sure I understand. In most cases the attributes are added to the > image (i.e. numpy array) when it is read. > >

2 1

Re: ndarray subclasses
by Kevin Keraudren Sept. 20, 2013

Sept. 20, 2013

Le 19/09/2013 10:36, Almar Klein a Ã¯Â¿Â½crit : > > 1) Allowing extra attributes for the Image class, like "sampling" that > specifies the distance between the pixels. This attribute can then be > used by algorithms to take anisotropy into account, and visualization > toolkits could use it to scale the image in the correct way > automatically. This may not seem a very common use case for 2D images, > but 3D data is usually not isotropic. > > Other attributes that I think may be of use for the Image class are > "origin" that specifies the location of the topleft pixel relative to > an arbitrary coordinate frame, and "meta" for the meta data (e.g. EXIF > tags). Hi, I am not related to the development of scikit-image but I guess its goal is to work on images in general and not get too specialised, for instance in Medical images. I am working on a Python interface for a C++ medical imaging library, and I chose to subclass np.ndarray, images keeping a header information ( 'dim', 'orientation', 'origin' and 'pixelSize'), the main feature is to keep track of those spatial coordinate when we slice or resample the array. As StÃ¯Â¿Â½fan pointed out " when you slice out a scalar, or sum, you get an Image object out!", but this has not been an issue so far, I just hide it by overriding: def __str__(self): if len(self.shape) == 0: return str(self.view(np.ndarray)) else: return self.__repr__() The documentation I placed online might give you more ideas on what could be done with a subclass of np.ndarray dedicated to images: http://www.doc.ic.ac.uk/~kpk09/irtk/#irtk.imread May I ask: if you were to add sampling information and spatial coordinates, such as an origin for the top-left pixel, how would you input that information into scikit-image, could you read it directly from the input files or would you need some user input? Kind regards, Kevin

2 1

Re: ndarray subclasses
by Stéfan van der Walt Sept. 19, 2013

Sept. 19, 2013

On Thu, Sep 19, 2013 at 12:03 PM, Almar Klein <almar.klein(a)gmail.com> wrote: > I suppose that the most important bit is that functions that support > anisotropy should look whether a "sampling" attribute is present in the > given array. I don't think we currently have any of these, but for now we can probably include a `sampling` argument to functions that support it. > I do not necessarily mean that a PointSet represents an image (although it > could), but more generally to for instance store the locations of detected > feature points. How do you think this would fit into the scope of image processing? (Asked out of curiosity, not at all to put the idea down.) Stéfan

1 0

Re: ndarray subclasses
by Almar Klein Sept. 19, 2013

Sept. 19, 2013

On 19 September 2013 11:41, Stéfan van der Walt <stefan(a)sun.ac.za> wrote: > Hi Almar > > On Thu, Sep 19, 2013 at 11:36 AM, Almar Klein <almar.klein(a)gmail.com> > wrote: > > Skimage also has the Image class (that inherits from np.ndarray), which I > > believe is mostly there to make the image be shown as an image in IPython > > notebook. What about taking this idea a bit further? In particular, I am > > thinking about two things: > > > > 1) Allowing extra attributes for the Image class, like "sampling" that > > specifies the distance between the pixels. This attribute can then be > used > > by algorithms to take anisotropy into account, and visualization toolkits > > could use it to scale the image in the correct way automatically. This > may > > not seem a very common use case for 2D images, but 3D data is usually not > > isotropic. > > We tried making this the default output of imread, but it caused some > problems (e.g. when you slice out a scalar, or sum, you get an Image > object out!), so we're back to vanilla arrays. I've had the same problem with my own "Image" class. You can solve this pretty well by overloading __getitem__ and __array_wrap__. A small example: https://gist.github.com/almarklein/6621363 That said, the Image > object is still quite useful. The Image class has a dictionary > container for adding any type of meta-data you want (e.g., origin, > sampling distance, etc.) so I don't know if it is necessary to > explicitly specify what those fields should be. > I suppose that the most important bit is that functions that support anisotropy should look whether a "sampling" attribute is present in the given array. > 2) Using a PointSet class to represent numpy arrays that are point sets or > > vector sets. Also such a class can make working with point data much > easier, > > both in internal algorithms, and for the end-user. > > > > An example class can be seen here: > > https://gist.github.com/almarklein/6620956 it allow things like > > appending/removing/inserting/ > popping individual points, and calculating > > things like normals, angles, distances, etc. > > In this case, I see an argument for having such a class, since it > behaves differently from a traditional image, and may need to be > resampled to a regular grid, etc. > I do not necessarily mean that a PointSet represents an image (although it could), but more generally to for instance store the locations of detected feature points. The computation of normals, angles, > etc. could be be utility functions outside the class (I guess I'm a > minimalist when it comes to class design). > You're probably be right about that. The most important bit is the manipulation of individual points inside the pointset.

1 0

Re: ndarray subclasses
by Stéfan van der Walt Sept. 19, 2013

Sept. 19, 2013

Hi Almar On Thu, Sep 19, 2013 at 11:36 AM, Almar Klein <almar.klein(a)gmail.com> wrote: > Skimage also has the Image class (that inherits from np.ndarray), which I > believe is mostly there to make the image be shown as an image in IPython > notebook. What about taking this idea a bit further? In particular, I am > thinking about two things: > > 1) Allowing extra attributes for the Image class, like "sampling" that > specifies the distance between the pixels. This attribute can then be used > by algorithms to take anisotropy into account, and visualization toolkits > could use it to scale the image in the correct way automatically. This may > not seem a very common use case for 2D images, but 3D data is usually not > isotropic. We tried making this the default output of imread, but it caused some problems (e.g. when you slice out a scalar, or sum, you get an Image object out!), so we're back to vanilla arrays. That said, the Image object is still quite useful. The Image class has a dictionary container for adding any type of meta-data you want (e.g., origin, sampling distance, etc.) so I don't know if it is necessary to explicitly specify what those fields should be. > 2) Using a PointSet class to represent numpy arrays that are point sets or > vector sets. Also such a class can make working with point data much easier, > both in internal algorithms, and for the end-user. > > An example class can be seen here: > https://gist.github.com/almarklein/6620956 it allow things like > appending/removing/inserting/popping individual points, and calculating > things like normals, angles, distances, etc. In this case, I see an argument for having such a class, since it behaves differently from a traditional image, and may need to be resampled to a regular grid, etc. The computation of normals, angles, etc. could be be utility functions outside the class (I guess I'm a minimalist when it comes to class design). Regards Stéfan

1 0

ndarray subclasses
by Almar Klein Sept. 19, 2013

Sept. 19, 2013

Hi, Many functions in skimage return a numpy array, but the meaning of that array can differ significantly: sometimes it represents an image, sometimes a set of locations or vectors, etc. Skimage also has the Image class (that inherits from np.ndarray), which I believe is mostly there to make the image be shown as an image in IPython notebook. What about taking this idea a bit further? In particular, I am thinking about two things: 1) Allowing extra attributes for the Image class, like "sampling" that specifies the distance between the pixels. This attribute can then be used by algorithms to take anisotropy into account, and visualization toolkits could use it to scale the image in the correct way automatically. This may not seem a very common use case for 2D images, but 3D data is usually not isotropic. Other attributes that I think may be of use for the Image class are "origin" that specifies the location of the topleft pixel relative to an arbitrary coordinate frame, and "meta" for the meta data (e.g. EXIF tags). 2) Using a PointSet class to represent numpy arrays that are point sets or vector sets. Also such a class can make working with point data much easier, both in internal algorithms, and for the end-user. An example class can be seen here: https://gist.github.com/almarklein/6620956 it allow things like appending/removing/inserting/popping individual points, and calculating things like normals, angles, distances, etc. I would like to stress that both Image and PointSet class do (and should) inherit from np.ndarray, such that they are completely compatible with existing code in skimage and other packages. I am interested to learn whether these ideas can find their way into skimage or not, Almar

1 0

Re: Keypoint matching and ORSA like algorithm
by Jérôme Kieffer Sept. 19, 2013

Sept. 19, 2013

Hello Séfan, On Wed, 18 Sep 2013 23:36:59 +0200 Stéfan van der Walt <stefan(a)sun.ac.za> wrote: > > Do you know an algorithm using those informations as well ? > > Using RANSAC (included in skimage as well), you can define any > criterium you choose for inliers. Wonderful, I will have a look. Cheers, -- Jérôme Kieffer <google(a)terre-adelie.org>

1 0

Keypoint matching and ORSA like algorithm
by Jérôme Kieffer Sept. 18, 2013

Sept. 18, 2013

Hi all, I am working on image alignment using SIFT algorithm; but this also applies to SURF. Sometimes one needs to "refine" matching because of a few wrong match in the list. The ORSA algorithm looks adapted (Moisan-Stival ORSA (Optimized Random Sampling Algorithm) algorithm to filter out the false matches using the epipolar geometry constraint) but it uses only the X and Y coordinate of the keypoint and not the angle + orientation. Maybe using those informations it would be easier to spot those outliers, no ? Do you know an algorithm using those informations as well ? Thanks, -- Jérôme Kieffer <google(a)terre-adelie.org>

2 1

Re: Any plans for 0.9 release?
by Pierre Villeneuve Sept. 14, 2013

Sept. 14, 2013

Stefan, Thanks for the reply. I was in fact hoping the solution would be this simple. Don't rush the release then just for me :) It looks like I'll be all set! Thanks again. On Friday, September 13, 2013 3:53:28 PM UTC-7, Stefan van der Walt wrote: > > Dear Pierre > > We are certainly due for a release, and while the next week is very > busy, I'm hoping to work on that in the two weeks following. I think > we're looking at around 15 October. In the mean time, you may use the > RANSAC tools without compiling by simply adding the appropriate Python > files into the correct locations. > > Regards > Stéfan > > On Sat, Sep 14, 2013 at 12:21 AM, Pierre Villeneuve > <pierre.v...(a)gmail.com <javascript:>> wrote: > > I am eagerly awaiting the new RANSAC model fitting tools in the current > dev > > version. I am working on Windows 8 x64 and it turns into a total mess > when > > I try to compile skimage myself. I normally rely on the binaries at > Chris > > Gohlke's site, but he only has the current stable version 0.8.2. I am > just > > curious to know if there is a planned timeframe for releasing 0.9? > Thanks! > > > > -- > > You received this message because you are subscribed to the Google > Groups > > "scikit-image" group. > > To unsubscribe from this group and stop receiving emails from it, send > an > > email to scikit-image...(a)googlegroups.com <javascript:>. > > For more options, visit https://groups.google.com/groups/opt_out. >

1 0