FAST, oFAST and rBRIEF #699

ankit-maverick · 2013-08-19T19:30:32Z

TO DO :

Replace the variables keypoints, scales, orientations, response with kpts_recarray:

    keypoints = np.array(np.dstack((row, col, octave, orientation, response)),
                         dtype=[('row', np.double), ('col', np.double),
                         ('octave', np.double), ('orientation', np.double),
                         ('response', np.double)])
    kpts_recarray = keypoints.view(np.recarray)

ankit-maverick · 2013-08-19T19:58:51Z

@ahojnnes : Please review fast.py and fast_cy.pyx.

ahojnnes · 2013-08-20T12:33:37Z

skimage/feature/fast.py

+
+
+def corner_fast(image, n=12, threshold=0.15):
+


remove empty line

ahojnnes · 2013-08-20T12:48:58Z

@ankit-maverick I really like your corner_fast implementation! I have not yet had time to review your orb function.

Two more notes to corner_fast. Please move the Python implementation to the existing corner.py and corner_cy.pyx files.

Have you done any comparisons to the OpenCV FAST results? Are they correct?

ahojnnes · 2013-08-20T12:50:14Z

skimage/feature/fast_cy.pyx

+                if circle_intensity > image[i, j] + threshold:
+                    # Brighter pixel
+                    bins[k] = 'b'
+                elif circle_intensity < image[i, j] - threshold:


You could save the value of image[i, j] in an intermediate variable.

ankit-maverick · 2013-08-20T13:52:38Z

@ahojnnes : I will do the comparison with OpenCV FAST and let you know the results. Before pushing, I verified it by picking 5-6 random corners detected by corner_fast and checked their surrounding circle intensities and they all followed the n and threshold. I now realize that my method would not account for the True negatives.

The ORB implementation currently is crude and just pure Python. I just wanted to code the part I had understood and hence the result. I will have to re-read the paper and check OpenCV's implementation again.

Remove image wrapper.

ahojnnes · 2013-08-21T20:02:40Z

skimage/feature/corner_cy.pyx

+cdef int[:] CP = (np.round(3 * np.cos(2 * np.pi * np.arange(16, dtype=np.double) / 16))).astype(np.int32)
+
+
+cdef inline _get_corner_response(double[:, ::1] image, int i, int j, char[:] bins, char check_state, int n, double threshold, double[:, ::1] corner_response):


Py_ssize_t for i, j, n

return response value, so you can avoid the corner_response parameter.

The function should get a different name, which associates it with "fast"

@ahojnnes : I considered avoiding the corner_response parameter, but realized that we would have to assign zero to all the non-corners using return 0.

What's wrong with returning 0?

Nothing wrong with it; just that I was thinking of some micro-optimization.

ahojnnes · 2013-08-21T20:10:30Z

Computing the angle of the keypoints should be a easy one now (I would implement it in Cython as well):

for n in range(keypoints.shape[0]):
keypoint_window = image[...]
moments = skimage.measure._moments.moments(...)
=> compute angle

Make sure to define the moments functions as cpdef so that they can be called efficiently from Cython, which also requires a pxd (header) file (have a look at the _shared package).

ahojnnes · 2013-08-21T20:43:49Z

skimage/feature/corner.py

+
+    image = img_as_float(image)
+    image = np.ascontiguousarray(image)
+    corner_response = _corner_fast(image, n, threshold)


To be consistent with the other corner detectors we should return the raw corner response. Non-maximal suppression is implemented in corner_peaks.

ahojnnes · 2013-08-22T12:54:41Z

skimage/feature/corner_cy.pyx

+            speed_sum_d = 0
+
+            for k in range(16):
+                circle_intensity = image[i + RP[k], j + CP[k]]


Does it make sense to save the intensity on the circle in an intermediate array double[16], so that you can reuse the values in _get_fast_corner_response, also we should consider to save lower_threshold = current_pixel - threshold and upper_threshold.

ahojnnes · 2013-10-25T12:40:48Z

@ankit-maverick Are you still working on this one?

ankit-maverick · 2013-10-25T12:59:36Z

@Ajohannes : I have some deadlines till Saturday. I will push on Sunday.

On Fri, Oct 25, 2013 at 6:10 PM, Johannes Schönberger <
notifications@github.com> wrote:

@ankit-maverick https://github.com/ankit-maverick Are you still working
on this one?

—
Reply to this email directly or view it on GitHubhttps://github.com//pull/699#issuecomment-27087572
.

ahojnnes · 2013-10-25T15:42:29Z

Great, ping when you’re ready.

ankit-maverick · 2013-10-28T21:23:33Z

@ahojnnes : Can you review the above changes? I will change the examples/tests/docs as soon as we are finalized on the implementation.

ahojnnes · 2013-10-30T14:17:00Z

skimage/feature/orb.py

+        return kpts_recarray
+    else:
+        best_indices = harris_measure.argsort()[::-1][:n_keypoints]
+        return kpts_recarray[best_indices]


You could also do kpts_recarray.sort(index=...)[:n_keypoints]

ahojnnes · 2013-10-30T14:22:12Z

@ankit-maverick Looks good to me. You could go ahead with the rest.

ankit-maverick · 2013-10-31T16:51:57Z

@ahojnnes : I think 'x' and 'y' is a better choice instead of 'row' and 'col'. Let me know your opinion.

ahojnnes · 2013-10-31T16:59:19Z

Hm, I always prefer row and col (r and c), because it makes crystal clear to which dimension of the array it belongs - whereas x and y has lead to several confusions in the past, also here at scikit-image…

But we could use both. x and y as an alias for row and col? What do you think?

ankit-maverick · 2013-10-31T17:08:18Z

@ahojnnes : Ok. I get your point now. x and y in our case would also imply that they are (0, 0) at left top which is ambiguous, as one might assume (0, 0) at left bottom. With row and col it is very clear that they start at the left top. So now I am against my old opinion.

ahojnnes · 2013-10-31T17:11:16Z

Great, let me know if you need any further help!

ankit-maverick · 2013-11-02T07:27:43Z

@ahojnnes , I feel that skimage.feature.util._create_keypoint_recarray should be made public so that we can let the users create a recarray from the corners returned by corner_peaks. Your opinion?

ahojnnes · 2013-11-02T11:17:21Z

Good point! +1 for making it public.

ankit-maverick · 2013-11-14T17:27:54Z

Hi @ahojnnes , apologies for stretching this so long. Can you review the minor changes in the previous 2-3 commits? I will update the doc/examples with the recarray changes now. Let me know if I am missing anything else.
Also, about the file path issue in orb_cy.pyx(https://groups.google.com/forum/#!topic/cython-users/1dJmqbMra0M), should I import the text file containing decision pairs in orb.py and then pass it onto `orb_cy.pyx? Or is there a better option?

ankit-maverick · 2013-11-15T07:14:09Z

Hi @ahojnnes , the matching performance seems to have reduced a bit after introducing the recarray changes. I guess it has something to do with float type for row and col. Please review the above changes.

ahojnnes · 2013-11-20T17:19:11Z

This looks good to me so far, but I have some more comments. I guess it saves time both for you and me if I do the changes directly in the code and send a PR against your branch and you can review them. What do you think?

ankit-maverick · 2013-11-20T19:52:13Z

@ahojnnes : Sure, I agree. I have my end-terms till the next Thursday, and then I can work on this and the FREAK descriptor. And woah!! This is the first time I am seeing your face!!

ankit-maverick · 2013-11-29T16:38:51Z

@ahojnnes : My end-terms are over and I am free. Can we decide how we should proceed on finishing this? In a few days or so, I am also planning to work on blob detection functions(https://en.wikipedia.org/wiki/Blob_detection).

ahojnnes · 2013-11-29T18:35:46Z

I’ll work on a PR this weekend and ping you here.

ahojnnes · 2014-01-23T13:22:00Z

@ankit-maverick I guess this PR can be closed then? Everything was contained in the ORB PR?

ankit-maverick · 2014-01-23T14:46:58Z

@ahojnnes : Ok. Closing this one :)

ankit-maverick added 4 commits August 16, 2013 15:01

Quick initial implementation of ORB descriptor in Python

810a1a2

Quick Python implementation of FAST corner detector

52e052e

Implemented FAST corner detector

d346744

Adding docstrings and commenting the code

7d76364

Adding high speed test for n>=12

bcf5bab

ahojnnes reviewed Aug 20, 2013
View reviewed changes

skimage/feature/fast.py

def corner_fast(image, n=12, threshold=0.15):

Copy link

Member

ahojnnes Aug 20, 2013

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

remove empty line

ahojnnes reviewed Aug 20, 2013
View reviewed changes

ankit-maverick and others added 7 commits August 21, 2013 01:19

Making various minor optimizations

9f0cba9

Transferring all the FAST code to corner.py and corner_cy.pyx

7cd5d12

Using inline function in corner_cy._corner_fast

395d4eb

Deleting FAST files

5ac298f

Converting image to float

9652989

Merge pull request scikit-image#702 from stefanv/remove_image_wrap

8d0f917

Remove image wrapper.

Adding additional parameter for setting shape for Non-max suppression

e993587

ahojnnes reviewed Aug 21, 2013
View reviewed changes

ankit-maverick added 2 commits August 22, 2013 04:06

Wrapping lines; returning corner_response at a pixel

f869d47

Returning the corner response

7100a82

ahojnnes reviewed Aug 22, 2013
View reviewed changes

Introducing recarray for storing keypoint variables

62eb8e1

ahojnnes reviewed Oct 30, 2013
View reviewed changes

Made recarray changes in docstrings and tests

c25f28c

ankit-maverick added 2 commits November 3, 2013 22:41

Extending recarray changes to BRIEF

7364594

Incorporating recarray changes to match.py

491de5b

ahojnnes mentioned this pull request Nov 12, 2013

Fix doctests and let TravisCI run doctests #811

Merged

Made recarray changes in the brief docstring

1461115

Recarray changes made in example

c635b6d

ahojnnes mentioned this pull request Nov 29, 2013

MRG: ORB #834

Merged

ankit-maverick closed this Jan 23, 2014

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

FAST, oFAST and rBRIEF #699

FAST, oFAST and rBRIEF #699

ankit-maverick commented Aug 19, 2013

ankit-maverick commented Aug 19, 2013

ahojnnes Aug 20, 2013

ahojnnes commented Aug 20, 2013

ahojnnes Aug 20, 2013

ankit-maverick commented Aug 20, 2013

ahojnnes Aug 21, 2013

ahojnnes Aug 21, 2013

ankit-maverick Aug 21, 2013

ahojnnes Aug 21, 2013

ankit-maverick Aug 21, 2013

ahojnnes commented Aug 21, 2013

ahojnnes Aug 21, 2013

ahojnnes Aug 22, 2013

ahojnnes commented Oct 25, 2013

ankit-maverick commented Oct 25, 2013

ahojnnes commented Oct 25, 2013

ankit-maverick commented Oct 28, 2013

ahojnnes Oct 30, 2013

ahojnnes commented Oct 30, 2013

ankit-maverick commented Oct 31, 2013

ahojnnes commented Oct 31, 2013

ankit-maverick commented Oct 31, 2013

ahojnnes commented Oct 31, 2013

ankit-maverick commented Nov 2, 2013

ahojnnes commented Nov 2, 2013

ankit-maverick commented Nov 14, 2013

ankit-maverick commented Nov 15, 2013

ahojnnes commented Nov 20, 2013

ankit-maverick commented Nov 20, 2013

ankit-maverick commented Nov 29, 2013

ahojnnes commented Nov 29, 2013

ahojnnes commented Jan 23, 2014

ankit-maverick commented Jan 23, 2014

		cdef int[:] CP = (np.round(3 * np.cos(2 * np.pi * np.arange(16, dtype=np.double) / 16))).astype(np.int32)


		cdef inline _get_corner_response(double[:, ::1] image, int i, int j, char[:] bins, char check_state, int n, double threshold, double[:, ::1] corner_response):

FAST, oFAST and rBRIEF #699

FAST, oFAST and rBRIEF #699

Conversation

ankit-maverick commented Aug 19, 2013

TO DO :

ankit-maverick commented Aug 19, 2013

Choose a reason for hiding this comment

ahojnnes commented Aug 20, 2013

Choose a reason for hiding this comment

ankit-maverick commented Aug 20, 2013

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ahojnnes commented Aug 21, 2013

Choose a reason for hiding this comment

Choose a reason for hiding this comment

ahojnnes commented Oct 25, 2013

ankit-maverick commented Oct 25, 2013

ahojnnes commented Oct 25, 2013

ankit-maverick commented Oct 28, 2013

Choose a reason for hiding this comment

ahojnnes commented Oct 30, 2013

ankit-maverick commented Oct 31, 2013

ahojnnes commented Oct 31, 2013

ankit-maverick commented Oct 31, 2013

ahojnnes commented Oct 31, 2013

ankit-maverick commented Nov 2, 2013

ahojnnes commented Nov 2, 2013

ankit-maverick commented Nov 14, 2013

ankit-maverick commented Nov 15, 2013

ahojnnes commented Nov 20, 2013

ankit-maverick commented Nov 20, 2013

ankit-maverick commented Nov 29, 2013

ahojnnes commented Nov 29, 2013

ahojnnes commented Jan 23, 2014

ankit-maverick commented Jan 23, 2014