mattnedrich / meanshift_py Goto Github PK

Hi mat,

Can you add some instructions about how to pass a 3D dimension image like RGB into the function? Besides, it looks when the size of image increase, you have to wait lots of time, so is there any recommended iteration number for breaking from the loop?

Best,

Haodong

Hi Matt!

First of all, thank you very much for this implementation, I found it very useful along with the blog post! :)

I did an implementation of a multivariate gaussian kernel, instead of a univariate gaussian kernel (as the one implemented). I needed for image processing, to find fixation points from raw gaze data.

I was wondering if you would be interested in a pull request with the implementation of the multivariate gaussian kernel.

Cheers,
Marc

PD: I wasn't sure where to put this messege, sorry if it shouldn't go in Issues ^^'

I transformed every RGB pixel of the image to a point in a 3d space, then mean shift the points to get the shifted points, but it took quite a long time to get the result because an image usually has hundreds times hundreds pixels.

The Gaussian Kernal algorithm in your code has a mistake:

def gaussian_kernel(distance, bandwidth):
    euclidean_distance = np.sqrt(((distance)**2).sum(axis=1))
    val = (1/(bandwidth*math.sqrt(2*math.pi))) * np.exp(-0.5*((euclidean_distance / bandwidth))**2)
    return 

Where val = (1/(bandwidth*math.sqrt(2*math.pi))) * np.exp(-0.5*((euclidean_distance / bandwidth))**2) should be
val = (1/(bandwidth*math.sqrt(2*math.pi))) * np.exp(-0.5*((euclidean_distance / bandwidth)**2)).

-0.5*((euclidean_distance / bandwidth))**2 -> -0.5*((euclidean_distance / bandwidth)**2)

Reference

• Radial_basis_function_kernel--wiki

I would like to generate images (like "ms_2d_bw_.8.gif"), but I don't see anything image related in the source code.
Where/how did you get those images?

Hi,

I don't use Python on regular basis, so maybe I don't understand something here, but how exactly am I supposed to include your code in my project? Usually I load stuff with PIP, but it seems that you haven't publish PIP package for this repo.

So I tried using git submodule and it seems ok, but I've got a problem with absolute imports - I use Python 3 so statements like import point_grouper as pg won't work unless I include your code in my project root directory - which would be so hacky. 😄

I think this could by fixed by replacing import lines with something like from . import point_grouper as pg. What do you think about this?

I feel that PIP package would be more elegant solution.

Hi mat, I appreciate your code! However in the following code, when I inspect the variable index, I found that the it is initialized once and then incremented in the loop, however it is never used to assign to anything. And this variable isn't passed to any other function, too. So I wonder is this a redundant variable? If so I think deleting it would make the logic clearer.

def group_points(self, points):
        group_assignment = []
        groups = []
        group_index = 0
        index = 0 # initialization
        for point in points:
            nearest_group_index = self._determine_nearest_group(point, groups)
            if nearest_group_index == None:
                # create new group
                groups.append([point])
                group_assignment.append(group_index)
                group_index += 1
            else:
                group_assignment.append(nearest_group_index)
                groups[nearest_group_index].append(point)
            index += 1 #increment
        return np.array(group_assignment)

I run the code with around 100 *10^3 points but I did not finish after 24 hours. I do not the reason.
Can any one help me to speed up the code with large size data.

Best,

hello, can you share the code for getting the image segment result ?

In the Attached screenshot , I had this error when i added my image to try it .

import numpy as np
import mean_shift as ms

mean_shifter = ms.MeanShift()

mean_shifter.cluster(np.array([[1,1], [1,2], [5,5]]), kernel_bandwidth = 2)
# no response, Ctrl + C

mean_shifter.cluster(np.array([[1.0,1.0], [1.0,2.0], [5.0,5.0]]), kernel_bandwidth = 2)
# works

Is it a bug? I would say it should accept integers also.