Convert COCO format segmentation annotation to LabelMe format

coco2labelme.py

import os
import json
import subprocess
import numpy as np
import pandas as pd
from skimage.measure import find_contours


class CocoDatasetHandler:
    def __init__(self, jsonpath, imgpath):
        with open(jsonpath, 'r') as jsonfile:
            ann = json.load(jsonfile)

        images = pd.DataFrame.from_dict(ann['images']).set_index('id')
        annotations = pd.DataFrame.from_dict(ann['annotations']).set_index('id')
        categories = pd.DataFrame.from_dict(ann['categories']).set_index('id')

        annotations = annotations.merge(images, left_on='image_id', right_index=True)
        annotations = annotations.merge(categories, left_on='category_id', right_index=True)
        annotations = annotations.assign(
            shapes=annotations.apply(self.coco2shape, axis=1))
        self.annotations = annotations
        self.labelme = {}

        self.imgpath = imgpath
        self.images = pd.DataFrame.from_dict(ann['images']).set_index('file_name')

    def coco2shape(self, row):
        if row.iscrowd == 1:
            shapes = self.rle2shape(row)
        elif row.iscrowd == 0:
            shapes = self.polygon2shape(row)
        return shapes

    def rle2shape(self, row):
        rle, shape = row['segmentation']['counts'], row['segmentation']['size']
        mask = self._rle_decode(rle, shape)
        padded_mask = np.zeros(
            (mask.shape[0]+2, mask.shape[1]+2),
            dtype=np.uint8,
        )
        padded_mask[1:-1, 1:-1] = mask
        points = find_contours(mask, 0.5)
        shapes = [
            [[int(point[1]), int(point[0])] for point in polygon]
            for polygon in points
        ]
        return shapes

    def _rle_decode(self, rle, shape):
        mask = np.zeros([shape[0] * shape[1]], np.bool)
        for idx, r in enumerate(rle):
            if idx < 1:
                s = 0
            else:
                s = sum(rle[:idx])
            e = s + r
            if e == s:
                continue
            assert 0 <= s < mask.shape[0]
            assert 1 <= e <= mask.shape[0], "shape: {}  s {}  e {} r {}".format(shape, s, e, r)
            if idx % 2 == 1:
                mask[s:e] = 1
        # Reshape and transpose
        mask = mask.reshape([shape[1], shape[0]]).T
        return mask

    def polygon2shape(self, row):
        # shapes: (n_polygons, n_points, 2)
        shapes = [
            [[int(points[2*i]), int(points[2*i+1])] for i in range(len(points)//2)]
            for points in row.segmentation
        ]
        return shapes

    def coco2labelme(self):
        fillColor = [255, 0, 0, 128]
        lineColor = [0, 255, 0, 128]

        groups = self.annotations.groupby('file_name')
        for file_idx, (filename, df) in enumerate(groups):
            record = {
                'imageData': None,
                'fillColor': fillColor,
                'lineColor': lineColor,
                'imagePath': filename,
                'imageHeight': int(self.images.loc[filename].height),
                'imageWidth': int(self.images.loc[filename].width),
            }
            record['shapes'] = []

            instance = {
                'line_color': None,
                'fill_color': None,
                'shape_type': "polygon",
            }
            for inst_idx, (_, row) in enumerate(df.iterrows()):
                for polygon in row.shapes:
                    copy_instance = instance.copy()
                    copy_instance.update({
                        'label': row['name'],
                        'group_id': inst_idx,
                        'points': polygon
                    })
                    record['shapes'].append(copy_instance)
            if filename not in self.labelme.keys():
                self.labelme[filename] = record

    def save_labelme(self, file_names, dirpath, save_json_only=False):
        if not os.path.exists(dirpath):
            os.makedirs(dirpath)
        else:
            raise ValueError(f"{dirpath} has existed")

        for file in file_names:
            filename = os.path.basename(os.path.splitext(file)[0])
            with open(os.path.join(dirpath, filename+'.json'), 'w') as jsonfile:
                json.dump(self.labelme[file], jsonfile, ensure_ascii=True, indent=2)
            if not save_json_only:
                subprocess.call(['cp', os.path.join(self.imgpath, file), dirpath])


ds = CocoDatasetHandler('cocodataset/annotations/instances_train2014.json', 'cocodataset/train2014/')
ds.coco2labelme()
ds.save_labelme(ds.labelme.keys(), 'cocodataset/labelme/train2014')

I am facing the same issue

My problem has been solved

Hi @travishsu,
Thank you for your quick reply!
Yes, my DataFile/train.json has an item with a 2-level file path on file_name. After changing this line of code you mentioned, it worked, but the problem is that it only split one JSON file from the train.json file and shown another error. Here is the snapshot of this error and result files.

I am very new in Python programming, so I am very sorry for asking this stupid question again.
This problem has been solved while I restart my computer and run this code again.

This problem has been solved while I restart my computer and run this code again.

Hey @TravisHu,
Perfectly , transferred the coco data set into a labelme format , but could not view in labelme, any idea?
Got this error everytime I try to load the image with labelme json files.

Hello @TuarAnup, sorry for the late reply.

I guess you were using save_json_only=True but didn't copy images to the same directory with output JSONs.
In Line-83 of this script, we set imageData to None, so we need a copy image placed with JSONs.

So if you were using save_json_only=True, try to copy images to the same directory with output JSONs.
Or try to set save_json_only=False.

Thx

Hello sir,
Thank you so much for your work!

I had this problem, the segmentation of one mask was converted into multiple segmentation.

for ex
COCO json file...
{
"id": 2,
"image_id": 1,
"segmentation": [
**[

]
[

]

]**
]
"iscrowd": 0,
"bbox": [
762.0,
496.0,
10.0,
10.0
],
"area": 83,
"category_id": 1394339
above segmentation has 3 masks of the same object.

When we convert it to label me
each of those masks is consider as separate segemntation

why so??

Hello @stphtan94117,

會是相對路徑不對嗎？我下載後有試成功

Hi @manaswakchaure,

Could you provide the entire value of the key "segmentation" in this example?
This script might call polygon2shape(L68) since "iscrowd" is 0.

@travishsu Thank you, sir, for your reply.

I have attached the files here https://drive.google.com/drive/folders/1cvPGxPGLCb-6fbDGVEHPDVX2bxPEvx3m?usp=sharing

Thank you!

This script might call polygon2shape(L68) since "iscrowd" is 0.

most of my images are crowded so should I fix it to 1? inorder to get the exact mask number?

Hi @manaswakchaure,

I found there are multiple lists nested in the value of segmentation so there'll be multiple converted masks for a single instance, and the converted masks will have the same group_id.

Besides, using iscrowd=0 is correct if the value of segmentation is in polygon format instead of RLE format.

Dear sir,

Thank you so much!

I got it. I made some modifications for getting those multiple masks as one single instance under one grup_id as needed. And verified it!

Thank you so much for your time.

不知道發這邊好不好
想請問作者有考慮新增png to json嗎?
也就是mask.png 轉成 coco.json (png2coco)
我原本參考這位作者的程式碼執行，他的資料集確實可以跑
但用我的資料去跑，產生COCO裡面的segmentatiom的點的座標，有些是負值，導致無法開起來
https://github.com/chrise96/image-to-coco-json-converter

我只有改動這邊，不曉得為何coco點座標會有負值

這是我的資料集檔案連結
https://drive.google.com/drive/folders/1butmjjGTgMIEr6bq1nQ3ejjm0M7oN_YR?usp=share_link
謝謝你

你好，我最後查出原因，只有把照片改成JPG即可
但有個極為困難的點，如果圖形是甜甜圈那種形狀，中間需要挖空
似乎coco json無法表示

用實際例子說明，假設有個正方形的農田，正中間有個農舍房子
因此農田polygon要正方形減去正中間農舍
但coco json的polygon的點，是輪廓組成，因此無法形成

不曉得你這邊有無辦法解決重疊的地方把它去除

Hi @stphtan94117,
我想 png2coco 可能不適合放在這個 coco2labelme 底下。

那位作者用的是多個 polygon 放在同一個 segmentation 且設定 iscrowd 為 0，
但如果有沒辦法用 polygon 表示的 instance，我想你可以考慮用將同一個 instance 的 mask 轉成 RLE format，且讓 iscrowd 設為 1

References

1. 找mask2rle
2. COCO Format