[PYTHON] How to create a label (mask) for segmentation with labelme (semantic segmentation mask)

↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓↓

Prerequisite knowledge

I am writing an article on the assumption that I have the following knowledge

• semantic segmentation (#semantic segmentation)
• mask（#label）

labelme is an app for creating labels

It's an app like this

• labelme is a label creation app made by python. You can label the image multiple times.
• There is no problem with other classes. So, of course, you can do it with 2 classes of semantic segmentation.
• Draw a line on the image with points and the result will be output as a JSON file
• If the JSON file is saved in the same folder as the image with the same name (it is saved with the same name by default), JSON will be loaded and opened in the annotated state when the image is loaded.
• The created points can be edited again.

Details on how to use labelme

Download label me

For windows

For mac

Label with label me

When you open labelme, the following window will open. (The image is a screenshot of the google search screen. It's hard to understand ...)

While expanding, draw a line at the point. It's very easy to zoom in and out.

I think it's about 1min with this proper feeling. Pretty easy to use. ..

Save the JSON from SAVE on the left. You can get data like this!

{
  "version": "4.5.6",
  "flags": {},
  "shapes": [
    {
      "label": "dog",
      "points": [
        [
          104.36893203883496,
          66.99029126213593
        ],
        [
          93.44660194174757,
          71.35922330097087
        ],
Omitted here
        [
          112.13592233009709,
          74.27184466019418
        ],
        [
          107.28155339805825,
          73.30097087378641
        ]
      ],
      "group_id": null,
      "shape_type": "polygon",
      "flags": {}
    }
  ],
  "imagePath": "Screenshot 2020-09-14 141400.png ",
  "imageData": "abridgement",
  "imageHeight": 405,
  "imageWidth": 535
}

Caution! Since this JSON point is a polygon point, it cannot be used as a mask as it is. .. So we need to convert this point to a mask.

Convert labelme file from JSON to MASK

At first I was thinking of implementing it myself (it's a hassle), but there was a function for it (that's right). labelme has a module called shape.py in the utils folder, and you can convert from JSON to MASK by using the shape_to_mask function.

However, I have to implement it subtly, so I will post the sample code.

import json
with open(path, "r",encoding="utf-8") as f:
    dj = json.load(f)
# dj['shapes'][0]Is for one label this time.
mask = shape_to_mask((dj['imageHeight'],dj['imageWidth']), dj['shapes'][0]['points'], shape_type=None,line_width=1, point_size=1)

mask_img = mask.astype(np.int)#boolean to 0,Convert to 1

#I'm using anaconda
import matplotlib.pyplot as plt
%matplotlib inline

plt.imshow(mask_img)

I was able to label it well. At this level, it takes about 2-3 minutes from opening the image to saving it, isn't it?

Finally

Please let me know if you have any recommended labeling tools other than label me. I wonder if some people have implemented it on their own.