VIA(VGG Image Annotator)マークアップデータをcoco formatに変換するコード

VGG Image Annotator(VIA)は、Visual Geometry Groupが開発したオープンソースの画像表示ツールです.長方形、円、楕円、多角形、点、線を表示できるオンラインおよびオフラインで使用できます.寸法が完了すると、csvとjsonファイル形式にエクスポートできます.
mmdetectionなどのフレームワークにVIAマークアップをサポートしていないデータがある場合は、変換コードを書く必要があります.
実は2つのjsonファイルの変換で、2つの寸法フォーマットが読み込まれた後に2つのdictフォーマットのデータで、私たちがしなければならないのは1層1層dictをダイヤルして、coco formatのdictの中に入れます.
まずフォーマットして2つのjsonファイルを出力することができて、比較的に異なって、1つのパッケージをインストールする必要があります

yum install -y jq

そして

cat a.json | jq

それから変換のコードを書くことができて、ここで私の書いたバージョンを提供して、親測して使いやすいです.中には無視できる小さな論理がある(自分で使ったら、変えない)

import json
import os
import cv2
import numpy as np

def _get_bbox(all_points_x, all_points_y):

	min_x , max_x = min(all_points_x), max(all_points_x)
	min_y , max_y = min(all_points_y), max(all_points_y)
	bbox = [min_x, min_y, max_x - min_x, max_y - min_y]
	return bbox

def _get_mask(all_points_x, all_points_y):
	min_x , max_x = min(all_points_x), max(all_points_x)
	min_y , max_y = min(all_points_y), max(all_points_y)
	for zz in range(len(all_points_x)): #find the biggest in list
		if(all_points_x[zz] == max_x):
			break
	left_lines = []
	right_lines = []
	mask = []
	for zzz in range(len(all_points_x)): # draw lines
		if zzz != zz :
			if all_points_y[zzz] < all_points_y[zz]:
				angle = abs(all_points_x[zzz] - all_points_x[zz])/abs(all_points_y[zzz] - all_points_y[zz] + 1e-7)
				#print('left',angle)
				left_lines.append([angle,all_points_x[zzz],all_points_y[zzz]])
			else:
				angle = abs(all_points_x[zzz] - all_points_x[zz])/abs(all_points_y[zzz] - all_points_y[zz] + 1e-7)
				#print('right',angle)
				right_lines.append([angle,all_points_x[zzz], all_points_y[zzz]])

	left_lines.sort()
	right_lines.sort(reverse=True)


	#print(left_lines)
	#print(right_lines)
	for line_num in range(len(left_lines)):
		mask.append(left_lines[line_num][1])
		mask.append(left_lines[line_num][2])
	for line_num in range(len(right_lines)):
		mask.append(right_lines[line_num][1])
		mask.append(right_lines[line_num][2])
	mask.append(all_points_x[zz])
	mask.append(all_points_y[zz])
	return [mask]

def _produce_category_id(data):   #       ，  ，index     id

	class_set = set()
	for i in data: 
		if(len(data[i]["regions"])!=0):
			if('name' in data[i]["regions"][0]["region_attributes"] and data[i]["regions"][0]["region_attributes"]["name"]!='USER'and data[i]["regions"][0]["region_attributes"]["name"]!='' and data[i]["regions"][0]["region_attributes"]["name"]!='ETTODAY' and data[i]["regions"][0]["region_attributes"]["name"]!='SICHUAN'and data[i]["regions"][0]["region_attributes"]["name"]!='RED' and data[i]["regions"][0]["region_attributes"]["name"]!='WEIXIN'):  #clean data
				class_set.add(data[i]["regions"][0]["region_attributes"]["name"].replace('
',''))
	class_set = list(class_set)
	class_set.sort()
	#print(class_set)
	#np.save('category_id_list.npy',class_set)
	return class_set


old_data = json.load(open('/Users/hank/Desktop/ad_12/annotations/via_region_data_train1.json')) #        
new_data = json.load(open('/Users/hank/Desktop/data/coco2017/annotations/instances_val2017.json')) #coco   

image_sample = new_data["images"][0]
annotations_sample = new_data["annotations"][0]
category_sample = new_data["categories"][0]

#print(annotations_sample)
idx = 0
ann_idx = 0
errorcount = 0
img_info = [] #image    
ann_info = [] #annotations    
cat_info = [] #category    
category_id_list = _produce_category_id(old_data)
image_path_list = []
#print(category_id_list)
user_count = 0
for _, image_i  in enumerate(old_data) :
	image_sample = {}
	image_name = old_data[image_i]["filename"]

	idx += 1
	image_path = os.path.join("/Users/hank/Desktop/ad_12/train/" , image_name)
	if(not os.path.exists(image_path)):   #        ，       
		errorcount += 1
		idx -= 1
		continue
	all_annotations_in_per = old_data[image_i]["regions"]
	flag = 0    #              label
	if(all_annotations_in_per ==[]):  #not tag
		image_path_list.append(image_path)
		print(image_path)
		os.remove(image_path)
	for iii in range(len(all_annotations_in_per)):
		if(not 'name' in old_data[image_i]["regions"][iii]["region_attributes"]):
			print(image_name)
		if ('name' in old_data[image_i]["regions"][iii]["region_attributes"]):
			#print(old_data[image_i]["regions"][iii]["region_attributes"]["name"])
			if(old_data[image_i]["regions"][iii]["region_attributes"]["name"] == 'USER' or old_data[image_i]["regions"][iii]["region_attributes"]["name"] == 'SICHUAN' or old_data[image_i]["regions"][iii]["region_attributes"]["name"] == 'ETTODAY' or old_data[image_i]["regions"][iii]["region_attributes"]["name"] == 'RED' or old_data[image_i]["regions"][iii]["region_attributes"]["name"] == 'WEIXIN'):
				flag = 1
				#print("happen")
	if flag == 2:
		#print(image_name)
		user_count += 1
		continue		
	if flag == 1:
		#user_count += 1
		os.remove(image_path)
		#print(image_path)
		continue
	
	img = cv2.imread(image_path) # image       
	if img is not None:
		h , w, _ = np.shape(img)
		image_sample["license"] = 2
		image_sample["coco_url"] = "http://images.cocodataset.org/val2017/000000500663.jpg"
		image_sample["file_name"] = image_name
		image_sample["id"] = idx
		image_sample["height"] = h
		image_sample["width"] = w
		image_sample["data_captured"] = "2013-11-17 21:48:19"
		image_sample["flickr_url"] = "http://farm1.staticflickr.com/198/488201322_ef2ebfeccb_z.jpg"

		img_info.append(image_sample)

	all_annotations_in_per = old_data[image_i]["regions"]   # annotation       
	#########change annotation to new type
	for j in range(len(all_annotations_in_per)):
		annotations_sample = {}
		#print(all_annotations_in_per)
		ann_idx += 1
		ann = all_annotations_in_per[j]   #dict[shape_attributes, region_attributes]
		ann_shape = ann["shape_attributes"]
		if ann_shape['name'] == 'polygon':
			all_points_x = ann_shape["all_points_x"]
			all_points_y = ann_shape["all_points_y"]
			bbox = _get_bbox(all_points_x, all_points_y)
			mask = _get_mask(all_points_x, all_points_y)
			annotations_sample['area'] = bbox[2] * bbox[3]
			annotations_sample['segmentation'] = mask
			annotations_sample['iscrowd'] = 0
			annotations_sample['image_id'] = idx
			annotations_sample['bbox'] = bbox
			if('name' in ann["region_attributes"]):
				#print(image_path)
				annotations_sample['category_id'] = category_id_list.index(ann["region_attributes"]["name"].replace('
','')) + 1
			else:
				continue
			annotations_sample['id'] = ann_idx
			#print(annotations_sample)
			ann_info.append(annotations_sample)

for k in range(len(category_id_list)):     # category       
	category_sample = {}
	category_sample['supercategory'] = category_id_list[k]
	category_sample['id'] = k + 1
	category_sample['name'] = category_id_list[k]
	print (category_sample)
	cat_info.append(category_sample)
#print(errorcount)
print('image——num',len(img_info))

new_data["annotations"] = ann_info
new_data["images"] = img_info
new_data["categories"] = cat_info
print(len(category_id_list))
#print(new_data["categories"])
new_data = json.dumps(new_data) #     json   
#print(category_id_list)
with open('/Users/hank/Desktop/ad_12/annotations/coco_train1.json','w') as json_file:
	json_file.write(new_data)
#print(type(new_data))