coco標註資訊與labelme標註資訊的詳解、相互轉換及視覺化
阿新 • • 發佈：2019-09-02
引言
在做例項分割或語義分割的時候，我們通常要用labelme進行標註，labelme標註的json檔案與coco資料集已經標註好的json檔案的格式和內容有差異。如果要用coco資料集的資訊，就要對json檔案進行修改和轉換。本部落格提供兩種格式的具體內容及含義以及兩種格式相互轉換的程式碼，並對兩種格式的json標註資訊進行視覺化。
1.coco格式的json標註資訊詳解及視覺化
從coco官網下載coco的資料集裡面，關於例項的標註資訊在“annotations_trainval2017.zip”壓縮檔案裡面的“instances_train2017.json”和“instances_val2017.json”裡面，分別是訓練集和驗證集的標註資訊。
下載地址：
訓練集圖片：http://images.cocodataset.org/zips/train2017.zip 驗證集圖片：http://images.cocodataset.org/zips/val2017.zip 測試集圖片：http://images.cocodataset.org/zips/test2017.zip 訓練集、驗證集標註資訊：http://images.cocodataset.org/annotations/annotations_trainval2017.zip http://images.cocodataset.org/annotations/stuff_annotations_trainval2017.zip http://images.cocodataset.org/annotations/panoptic_annotations_trainval2017.zip
由於“instances_train2017.json”裡面把所有訓練集圖片的標註資訊整合到一個檔案了，檔案非常大，不太好檢視內部具體內容。我從這個檔案中提取出一張圖片的資訊儲存成一個新的json檔案。
json檔案內主要欄位：
程式碼如下：
# -*- coding: utf-8 -*-
from __future__ import print_function
"""
Created on Thu Aug 15 13:58:40 2019
@author: Taoting
"""
'''
從coco的標註檔案裡提取一張圖片對應的json資訊，並儲存成新的json檔案（以instance為例，其他的類似）。
'''
import matplotlib.pyplot as plt
import os, sys, zipfile
import urllib.request
import shutil
import numpy as np
import skimage.io as io
import pylab
import json
from pycocotools.coco import COCO
pylab.rcParams['figure.figsize'] = (8.0, 10.0)
json_file='../../../coco dataset/annotations_trainval2017/instances_val2017.json' # # Object Instance 型別的標註
# json_file='./annotations/person_keypoints_val2017.json' # Object Keypoint 型別的標註格式
# json_file='./annotations/captions_val2017.json' # Image Caption的標註格式
data=json.load(open(json_file,'r'))
data_2={}
data_2['info']=data['info']
data_2['licenses']=data['licenses']
data_2['images']=[data['images'][0]] # 只提取第一張圖片
data_2['categories']=data['categories']
annotation=[]
# 通過imgID 找到其所有instance
imgID=data_2['images'][0]['id']
#print(imgID)#397133
# initialize COCO api for instance annotations
coco=COCO(json_file)
img = coco.loadImgs([imgID])
#print(img)#[{'license': 4, 'file_name': '000000397133.jpg', 'coco_url': 'http://images.cocodataset.org/val2017/000000397133.jpg',
# 'height': 427, 'width': 640, 'date_captured': '2013-11-14 17:02:52', 'flickr_url': 'http://farm7.staticflickr.com/6116/6255196340_da26cf2c9e_z.jpg', 'id': 397133}]
#print(img['file_name'])
# load and display image
I = io.imread('../../../coco dataset/val2017/%s' % img[0]['file_name'])
# use url to load image
#I = io.imread(img['coco_url'])
#plt.axis('off')
#plt.imshow(I)
#plt.show()
for ann in data['annotations']:
if ann['image_id']==imgID:
annotation.append(ann)
data_2['annotations']=annotation
# 儲存到新的json
json.dump(data_2,open('./{}.json'.format(str(img[0]['file_name']).split('.')[0]),'w'),indent=4)
得到一張圖片的標註資訊如下，包含5大部分的欄位資訊。
"info"的value是一個dict，儲存資料集的一些基本資訊，我們不需要關注；
"licenses"的value是一個list，儲存license資訊，我們不需要關注；
"categories"的value是一個list，儲存資料集的類別資訊，包括類別的超類、類別id、類別名稱；
“images”的value是一個list，儲存這張圖片的基本資訊，包括圖片名、長、寬、id等重要資訊；
"annotations"的value是一個list，儲存這張圖片的標註資訊，非常重要，list中的每一個元素是一個dict，也即一個標註物件（instance）的資訊。包括的欄位有"segmentation"：標註點的座標，從第一個的x,y座標一直到最後一個點的x,y座標；"area"是標註的閉合多邊形的面積； "iscrowd"表示物件之間是否有重疊；"image_id"是圖片的id；“bbox”是instance的邊界框的左上角的x,y，邊界框的寬和高；"category_id"是這個instance對應的類別id；"id"表示此instance標註資訊在所有instance標註資訊中的id。
{
"info": {
"description": "COCO 2017 Dataset",
"url": "http://cocodataset.org",
"version": "1.0",
"year": 2017,
"contributor": "COCO Consortium",
"date_created": "2017/09/01"
},
"licenses": [
{
"url": "http://creativecommons.org/licenses/by-nc-sa/2.0/",
"id": 1,
"name": "Attribution-NonCommercial-ShareAlike License"
},
{
"url": "http://creativecommons.org/licenses/by-nc/2.0/",
"id": 2,
"name": "Attribution-NonCommercial License"
},
...(太長，省略)
],
"categories": [
{
"supercategory": "person",
"id": 1,
"name": "person"
},
...(太長，省略)
],
"images": [
{
"license": 2,
"file_name": "000000000049.jpg",
"coco_url": "http://images.cocodataset.org/train2017/000000000049.jpg",
"height": 500,
"width": 381,
"date_captured": "2013-11-14 20:00:23",
"flickr_url": "http://farm4.staticflickr.com/3250/2883102207_bcba5527a7_z.jpg",
"id": 49
}
],
"annotations": [
{
"segmentation": [
[
181.59,
363.43,
...(太長，省略)
]
],
"area": 8451.22405,
"iscrowd": 0,
"image_id": 49,
"bbox": [
162.57,
226.56,
130.41,
184.43
],
"category_id": 19,
"id": 56407
},
...(太長，省略)
]
}
我們對這個新coco格式的json檔案進行視覺化：
# -*- coding: utf-8 -*-
"""
Created on Thu Aug 15 14:48:12 2019
@author: Taoting
"""
from __future__ import print_function
import matplotlib.pyplot as plt
# ~ from pycocotools.coco import COCO
from coco import COCO
import os, sys, zipfile
import urllib.request
import shutil
import numpy as np
import skimage.io as io
import pylab
pylab.rcParams['figure.figsize'] = (8.0, 10.0)
annFile='./modified_satisfied_json_train2017/000000000149.json'#json檔案路徑
coco=COCO(annFile)
cats = coco.loadCats(coco.getCatIds())
nms=[cat['name'] for cat in cats]
nms = set([cat['supercategory'] for cat in cats])
imgIds = coco.getImgIds()
img = coco.loadImgs(imgIds[0])[0]
dataType = './satisfied_images_train2017'
I = io.imread('%s/%s'%(dataType,img['file_name']))
plt.axis('off')
plt.imshow(I)
plt.show()
# 載入和視覺化instance標註資訊
catIds=[]
for ann in coco.dataset['annotations']:
if ann['image_id']==imgIds[0]:
catIds.append(ann['category_id'])
plt.imshow(I); plt.axis('off')
annIds = coco.getAnnIds(imgIds=img['id'], catIds=catIds, iscrowd=None)
anns = coco.loadAnns(annIds)
coco.showAnns(anns)
plt.show()
coco格式的json檔案視覺化instance的mask
視覺化結果如下：
2.labelme格式的json標註資訊詳解及視覺化
labelme標註工具標註的json格式與coco的格式有差異：
重點關注的是：
"shapes"：儲存標註instance的閉合多邊形的資訊，重點關注：label：類別名稱；points：閉合多邊形的每個點的x,y座標；
"line_color"：閉合多邊形的邊界線顏色；
"fill_color"：閉合多邊形的填充顏色；
"imagePath"：圖片名稱；
"imageData"：圖片路徑（加密後）；
"imageHeight"：圖片高；
"imageWidth"：圖片寬；
利用labelme提供的介面將標註好的json進行視覺化程式碼：
# -*- coding:utf-8 -*-
"""
@author: Taoting
將用labeime標註格式的json進行視覺化
"""
import json
import matplotlib.pyplot as plt
import skimage.io as io
from labelme import utils
def main():
json_path = './PATH/TO/JSON'
data = json.load(open(json_path))
img = io.imread('%s/%s'%('./PATH/TO/IMAGE',data['imagePath']))
lab, lab_names = utils.labelme_shapes_to_label(img.shape, data['shapes'])
captions = ['%d: %s' % (l, name) for l, name in enumerate(lab_names)]
lab_ok = utils.draw_label(lab, img, captions)
plt.subplot(121)
plt.imshow(img)
plt.subplot(122)
plt.imshow(lab_ok)
plt.show()
if __name__ == '__main__':
main()
視覺化結果：
可以看到右圖中的mask的視覺化效果
結合1和2中的兩種格式的json，我們只需要針對格式的差異對json檔案做修改，就能將格式進行互相轉換。
3.coco格式的json轉labelme格式的json
直接上程式碼：
# -*- coding: utf-8 -*-
"""
@author: Taoting
將用coco格式的json轉化成labeime標註格式的json
"""
import json
import cv2
import numpy as np
import os
#用一個labelme格式的json作為參考，因為很多資訊都是相同的，不需要修改。
def reference_labelme_json():
ref_json_path = 'reference_labelme.json'
data=json.load(open(ref_json_path))
return data
def labelme_shapes(data,data_ref):
shapes = []
label_num = {'person':0,'bicycle':0,'car':0,'motorcycle':0,'bus':0,'train':0,'truck':0}#根據你的資料來修改
for ann in data['annotations']:
shape = {}
class_name = [i['name'] for i in data['categories'] if i['id'] == ann['category_id']]
#label要對應每一類從_1開始編號
label_num[class_name[0]] += 1
shape['label'] = class_name[0] + '_' + str(label_num[class_name[0]])
shape['line_color'] = data_ref['shapes'][0]['line_color']
shape['fill_color'] = data_ref['shapes'][0]['fill_color']
shape['points'] = []
# ~ print(ann['segmentation'])
if not type(ann['segmentation']) == list:
continue
else:
x = ann['segmentation'][0][::2]#奇數個是x的座標
y = ann['segmentation'][0][1::2]#偶數個是y的座標
for j in range(len(x)):
shape['points'].append([x[j], y[j]])
shape['shape_type'] = data_ref['shapes'][0]['shape_type']
shape['flags'] = data_ref['shapes'][0]['flags']
shapes.append(shape)
return shapes
def Coco2labelme(json_path,data_ref):
with open(json_path,'r') as fp:
data = json.load(fp) # 載入json檔案
data_labelme={}
data_labelme['version'] = data_ref['version']
data_labelme['flags'] = data_ref['flags']
data_labelme['shapes'] = labelme_shapes(data,data_ref)
data_labelme['lineColor'] = data_ref['lineColor']
data_labelme['fillColor'] = data_ref['fillColor']
data_labelme['imagePath'] = data['images'][0]['file_name']
data_labelme['imageData'] = None
# ~ data_labelme['imageData'] = data_ref['imageData']
data_labelme['imageHeight'] = data['images'][0]['height']
data_labelme['imageWidth'] = data['images'][0]['width']
return data_labelme
if __name__ == '__main__':
root_dir = './ROOT DIR'
json_list = os.listdir(root_dir)
#參考的json
data_ref = reference_labelme_json()
for json_path in json_list:
if json_path.split('.')[-1] == 'json':
print('當前檔案： ', json_path)
data_labelme= Coco2labelme(os.path.join(root_dir,json_path), data_ref)
file_name = data_labelme['imagePath']
# 儲存json檔案
json.dump(data_labelme,open('./PATH/%s.json' % file_name.split('.')[0],'w'),indent=4)
用2中的視覺化程式碼檢驗是否正確轉換。
4.labelme格式的json轉coco格式的json
直接上程式碼：
# -*- coding: utf-8 -*-
"""Created on Thu Aug 15 15:05:56 2019
@author: Taoting
將用labeime標註的json轉化成coco格式的json
"""
import json
import cv2
import numpy as np
import os
#用閉包實現計數器
def counter():
cnt = 1000000
def increce():
nonlocal cnt
x = cnt
cnt += 1
return x
return increce
def p_images(data,data_coco):
images=[]
image={}
file_name=data['imagePath'].split('\\')[-1]
image['file_name']=file_name
image['id']=int(file_name.split('.')[0])
image['height']=data['imageHeight']
image['width']=data['imageWidth']
img=None
images.append(image)
data_coco['images']=images
return file_name
#用一個coco格式的json做參考
def modify_categories():
ref_json_path = 'reference.json'
data=json.load(open(ref_json_path))
modified_categories = []
catNms=['person','bicycle','car','motorcycle','truck','bus']#根據你的資料修改
for i,cat in enumerate(data['categories']):
if cat['name'] in catNms:
modified_categories.append(cat)
else:
pass
return modified_categories,data['info'],data['licenses']
def p_annotation(data,data_coco,cnt):
# annotations
annotations=[]
for i in range(len(data['shapes'])):
annotation={}
annotation['segmentation']=[list(np.asarray(data['shapes'][i]['points']).flatten())] # data['shapes'][0]['points']
annotation['iscrowd']=0
annotation['image_id']=data_coco['images'][0]['id']
#找出標註點中的外接矩形的四個點
x = annotation['segmentation'][0][::2]#奇數個是x的座標
y = annotation['segmentation'][0][1::2]#偶數個是y的座標
print(x,y)
x_left = min(x)-1#往外擴充套件1個畫素，也可以不擴充套件
y_left = min(y)-1
w = max(x) - min(x)+1
h = max(y) - min(y)+1
annotation['bbox']=[x_left,y_left,w,h] # [左上角x,y以及寬和高]
cat_list_dict = [cat for cat in data_coco['categories'] if cat['name'] == data['shapes'][i]['label'].split('_')[1]]#注意這裡是跟標註時填類別的方式有關
annotation['category_id']=cat_list_dict[0]['id']
annotation['id'] = cnt() # 第一個物件 這個ID也不能重複，如果下一張圖，id不能取1，需從1 開始往下取
#print('cnt', annotation['id'])
#print('annotation',annotation)
annotations.append(annotation)
#print('annotations',annotations)
data_coco['annotations']=annotations
#print(data_coco['annotations'])
#return data_coco
def Labelme2coco(json_path,cnt):
with open(json_path,'r') as fp:
data = json.load(fp) # 載入json檔案
data_coco={}
# images
file_name = p_images(data,data_coco)
# categories
modified_categories, info, p_license = modify_categories()
data_coco['categories'] = modified_categories
#print(data_coco['categories'])
# info
data_coco['info'] = info
# license
data_coco['license'] = p_license
# annotations
p_annotation(data,data_coco,cnt)
#print(data_coco['annotations'])
return data_coco,file_name
if __name__ == '__main__':
root_dir = './ROOT DIR'
json_list = os.listdir(root_dir)
cnt = counter()
for json_path in json_list:
if json_path.split('.')[-1] == 'json':
data_coco,file_name = Labelme2coco(os.path.join(root_dir,json_path),cnt)
# 儲存json檔案
json.dump(data_coco,open('./PATH/%s.json' % file_name.split('.')[0],'w'),indent=4)
用1中的視覺化程式碼檢驗是否正確轉