is there a way to get UCM - segmentation using python and opencv ? about synthtext

[kaislaribi]

@sszzy0317 here is it https://www.codepile.net/pile/N8akd162

[ankush-me]

You can use any image segmentation algorithm. Skimage has several:
http://scikit-image.org/docs/dev/api/skimage.segmentation.html

[kaislar]

thanks, i got the segmentation. Then, in which format do you store the segmented image in the .mat file
I tried as a numpy array and i got this error when i try to execute the foodfill.py script
`
Traceback (most recent call last):
File "floodFill.py", line 132, in
process_db_parallel(base_dir)
File "floodFill.py", line 115, in process_db_parallel
for res in ucm_result:
File "/home/deep/miniconda3/envs/env_dataset/lib/python2.7/multiprocessing/pool.py", line 668, in next
raise value
AttributeError: 'numpy.int64' object has no attribute 'encode'

`

[ankush-me]

Please check the format details in dset.h5

[kaislar]

yes, i got the segmentation and depth and added them to the dataset
i got an error :

File "/media/sf_PartageVM/Dataset/SynthText/synthgen.py", line 619, in render_text
regions = self.filter_for_placement(xyz,seg,regions)
File "/media/sf_PartageVM/Dataset/SynthText/synthgen.py", line 395, in filter_for_placement
res = get_text_placement_mask(xyz,seg==l,regions['coeff'][idx],pad=2)
File "/media/sf_PartageVM/Dataset/SynthText/synthgen.py", line 252, in get_text_placement_mask
s = rescale_frontoparallel(pts_tmp,boxR,pts[0])
File "/media/sf_PartageVM/Dataset/SynthText/synthgen.py", line 200, in rescale_frontoparallel
lt1 = np.linalg.norm(p_im[n1,:]-p_im[n0,:])
IndexError: too many indices for array

Apparently it is not detecting contours !
Here is the dataset with one new image https://ufile.io/u2fs3
if you have any idea tell me...
thank you.

[junedgar]

@kaislar Did you get ucm.mat by using python? I got the data of segmentation, but there may someting wrong with the format. I ran with the error below:

  File "/Users/edgar/project/pyCodes/SynthText/synthgen.py", line 626, in render_text
    regions = self.filter_for_placement(xyz,seg,regions)
  File "/Users/edgar/project/pyCodes/SynthText/synthgen.py", line 395, in filter_for_placement
    res = get_text_placement_mask(xyz,seg==l,regions['coeff'][idx],pad=2)
  File "/Users/edgar/project/pyCodes/SynthText/synthgen.py", line 243, in get_text_placement_mask
    R2d = su.unrotate2d(box.copy())
  File "/Users/edgar/project/pyCodes/SynthText/synth_utils.py", line 262, in unrotate2d
    l,R = np.linalg.eig(pts.T.dot(pts))
  File "/Users/edgar/.pyenv/versions/anaconda2-4.0.0/lib/python2.7/site-packages/numpy/linalg/linalg.py", line 1122, in eig
    _assertFinite(a)
  File "/Users/edgar/.pyenv/versions/anaconda2-4.0.0/lib/python2.7/site-packages/numpy/linalg/linalg.py", line 217, in _assertFinite
    raise LinAlgError("Array must not contain infs or NaNs")
LinAlgError: Array must not contain infs or NaNs

If you had solve the problem, can you share me the idea?

[kaislar]

i wrote a python script using slic from skimage.segmentation to perform segmentation and then i count the area of each segmented region.
So i didn't use the .mat file i created my own dataset with new images

[junedgar]

@kaislar
thanks for your reply .
I tried the skimage.segmentation method, but got the some error. Maybe there is something wrong with other data. I generated the depth image using the tensorflow branch of FCRN.

[kaislar]

For segmented image did you get the labels and areas attributes stored the database. Look to the sample database to understand how to store them

[sszzy0317]

@kaislar
would you share your python file for segmentation, I have trouble with using the OpenCV

[sszzy0317]

@kaislar THANKS, I Have built the model like
"segment = slic(img, n_segments=100, compactness=20)"
but I didn't known the meaning of the parameter of n_segments and compactness

[mohammedayub44]

@kaislaribi Thanks for sharing your code snippet. It was helpful.

@ankush-me I managed to run create a dset.h5 file using custom images with homes. It returns synthetic text as expected for some images (like shown below). However for some images it is throwing error similar to what is mentioned by @kaislaribi

Stack Trace of Errors (for some of the images) -

Traceback (most recent call last):
  File "/media/mayub/OS/Users/Admin/Documents/GitHub/SynthText/synthgen.py", line 615, in render_text
    regions = self.filter_for_placement(xyz,seg,regions)
  File "/media/mayub/OS/Users/Admin/Documents/GitHub/SynthText/synthgen.py", line 391, in filter_for_placement
    res = get_text_placement_mask(xyz,seg==l,regions['coeff'][idx],pad=2)
  File "/media/mayub/OS/Users/Admin/Documents/GitHub/SynthText/synthgen.py", line 250, in get_text_placement_mask
    s = rescale_frontoparallel(pts_tmp,boxR,pts[0])
  File "/media/mayub/OS/Users/Admin/Documents/GitHub/SynthText/synthgen.py", line 200, in rescale_frontoparallel
    lt1 = np.linalg.norm(p_im[n1,:]-p_im[n0,:])
IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed

I'm guessing this is because it did not find any good segments ? I might be wrong.
Here is one image it threw error on.

For reference, this is how the masks and segments look like for the error image:

Any specific settings I can change to accommodate these images.

Thanks in advance !