Plygon is not connected as a contour

[MingxuanGu]

Plygon is not connected as a contour

[ChristianMarzahl]

Dear StormyaP,

is "closed" set in your settings?

[MingxuanGu]

Hi Christian, Yes it’s in my setting. Best, Mingxuan From: Christian Marzahl ***@***.***> Sent: Tuesday, November 16, 2021 4:05 PM To: ChristianMarzahl/Exact ***@***.***> Cc: Mingxuan Gu ***@***.***>; Author ***@***.***> Subject: Re: [ChristianMarzahl/Exact] Plygon is not connected as a contour (Issue #73) Dear StormyaP, is "closed" set in your settings? <https://user-images.githubusercontent.com/22664224/142010070-d036438a-0e9e-4c2d-a3bc-70dfd2722286.png> — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub <#73 (comment)> , or unsubscribe <https://github.com/notifications/unsubscribe-auth/AIEWKKR2RU5QSUHN2C4CBUDUMJXINANCNFSM5IEO6XUQ> . Triage notifications on the go with GitHub Mobile for iOS <https://apps.apple.com/app/apple-store/id1477376905?ct=notification-email&mt=8&pt=524675> or Android <https://play.google.com/store/apps/details?id=com.github.android&referrer=utm_campaign%3Dnotification-email%26utm_medium%3Demail%26utm_source%3Dgithub> . <https://github.com/notifications/beacon/AIEWKKTXIDNKACP5XVZKRA3UMJXINA5CNFSM5IEO6XU2YY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOHHLIWRA.gif>

[MingxuanGu]

this is my setting

[ChristianMarzahl]

@OettlM do you have an idea?

With kind regards,
Christian

[OettlM]

@MingxuanGu can you share the way the second image is created? How to you convert the polygon vector to pixel wise segmentations?

[MingxuanGu]

I first exported the annotation.

Then I read in the .txt file with

csv_file = pd.read_csv(annotation_dir, sep='|', header=None, names=columns, index_col=False)

I created a label array to be filled with labels

anno = np.zeros((20, 2038, 1996))

Then I loop over the "csv_file" and get the vector, class name, and frame number of each row in the "csv_file"

for i in range(len(csv_file)):
    vector_dir = csv_file.loc[[i]].vector[i]
    vector_dir = ast.literal_eval(vector_dir)
    class_name = csv_file.loc[[i]].class_name[i]
    class_num = 1 if class_name == 'Blood_vessel_polygon' else 2
    frame_num = csv_file.loc[[i]].frame[i]
    j = 0
    while 1:
        try:
            coordx = vector_dir['x' + str(j)]
            coordy = vector_dir['y' + str(j)]
            anno[frame_num - 1, coordy, coordx] = class_num
        except KeyError:
            break
        j += 1

After that, I fill the holes for each object in both class:

from scipy.ndimage.morphology import binary_fill_holes
anno1 = np.where(anno == 1, 1, 0)
anno2 = np.where(anno == 2, 1, 0)
for i in range(len(anno1)):
    anno1[i] = binary_fill_holes(anno1[i])
    anno2[i] = binary_fill_holes(anno2[i])
anno = np.where(anno1 == 1, 1, anno)
anno = np.where(anno2 == 1, 2, anno)

At the end, I get this "anno" as the label.

[OettlM]

Okay, I think I see where the issue is here.
The data you get is a vector containing each point that is part of the polygon annotation.

What you do in your code is to assign each pixel the value of the point it contains. However, the annotations in EXACT are polygons, so you dont nessesarly have a point for every pixel of the contour, since it just connects each point to the next.

A solution to this behaviour would be to use a function that interprets the vector as a polygon and draws it into your image.
OpenCV offers functionality to easily do that. https://docs.opencv.org/3.4/d4/d73/tutorial_py_contours_begin.html

Here is an example how you can use cv2.drawContours to draw the polygons. You can ignore the middle part where I sort the annotations, depending on the area size, which is only relevant if you have overlapping annotations.

In this case the function is also parameterized to directly draw filled polygons, so you dont have to do that yourself afterwards.

[MingxuanGu]

Amazing, it works! Thanks!

[ChristianMarzahl]

Big thanks to @OettlM