NPY/PNG to Json annotation

[paresman]

Hello !
I had an issue for converting back png / npy to json, and the code I saw in the issue
#165
and
#374
Did not helped me at all, specifically for image where border were labelised, in which case skimage has issue, and case with many labels and instance on image, in which case
#165 was not helpfull.
So i coded this in a couple of hour for solving the problem
I hope it could help someone

from skimage import measure
import labelme
import numpy as np
from PIL import Image
import os
import json
null = None
import glob
import os, sys
sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from labels import labels_list
# Example of labels_list :
# labels_list = [
#     "_background_",
#     "dog",
#     "cat",
# ]

files = glob.glob("*.npy")

# For each file :
for file_name in files :
    file_name = file_name.replace(".npy","")
    # Open raw image for computing imagedata used in json
    img = Image.open(os.path.join("..","..","img_raw",file_name+".png"))
    mask = np.asarray(img)
    imageData = labelme.utils.img_arr_to_b64(mask)
    # Open generated npy
    mask = np.load(file_name+".npy")
    # This one is maybe wrong, don't know i work with square images
    imageHeight = mask.shape[0]
    imageWidth = mask.shape[1]
    # One annotation per file
    annotation = {
        "imageHeight": imageHeight,
        "imageWidth": imageWidth,
        "shapes": [],
        "imagePath": file_name,
        "imageData": imageData.decode('utf-8'),
        "version": "5.0.1",
        "flags": {},
    }
    # Set background for border for a better recognition of shape
    mask[0,:]=0
    mask[:,0]=0
    mask[imageHeight-1,:]=0
    mask[:,imageWidth-1]=0
    # Search for each label if they exist in file
    for label, label_string in enumerate(labels_list):
        # Build mask_label to numpy
        mask_label = np.where(mask == label,255,0.0)
        contours = measure.find_contours(mask_label)
        # contours = sorted(contours, key=len, reverse=True)[:1]
        # for n, contour in enumerate(contours):
        for contour in contours:
            coords = measure.approximate_polygon(contour, tolerance=3)[:-1]
            segmentation = np.flip(coords, axis=1).tolist()
            # segmentation = coords.tolist()
            polygon = {
                "label": label_string,
                "group_id": null,
                "shape_type": "polygon",
                "flags": {},
                "points": segmentation,
            }
            annotation["shapes"].append(polygon)
    # When all labels have been searched, we can create the json
    with open(file_name+".json", 'w') as outfile:
        json.dump(annotation, outfile, indent=2)