实验数据集介绍
本次使用的数据集是NWPU VHR-10数据集，涉及10个目标，一共有650张图像，650张图像分别对应了650个标注文件，标注文件中一行为标注一个目标，其中前面两个值分别表示左上角和右下角的坐标位置，最后一个值表示目标的类别，以下是10个目标对应的ID。
1-airplane, 2-ship, 3-storage tank, 4-baseball diamond, 5-tennis court, 6-basketball court, 7-ground track field, 8-harbor, 9-bridge, 10-vehicle
现在要在主流的一些目标检测框架上进行训练和测试，首先要转化为主流的目标检测的格式，包括coco格式、voc格式和yolo格式，方便在这些网络的基础上进行微调。
YOLO格式
yolo的格式也分为多种：
yolov3格式版本：
代码地址：eriklindernoren/PyTorch-YOLOv3: Minimal PyTorch implementation of YOLOv3 ()
格式要求：Move your annotations to
data/custom/labels/. The dataloader expects that the annotation file corresponding to the imagedata/custom/images/train.jpghas the pathdata/custom/labels/train.txt. Each row in the annotation file should define one bounding box, using the syntaxlabel_idx x_center y_center width height. The coordinates should be scaled[0, 1], and thelabel_idxshould be zero-indexed and correspond to the row number of the class name indata/custom/classes.names.每张图片对应一个标注文件，其中每一行表示一个目标，从左向右依次是
label_idx x_center y_center width height，label_idx是从0开始的。数据格式：类别id 中心点x坐标 中心点y坐标 w h
常用的数据格式就是这样，关键在于如何将代码进行转化，主要是左上角坐标和右下角坐标到yolo这四个格式之间的转化公式。
# 标准的转化形式如下：x, y, x2, y2 = obj['bbox']class_name = obj['name']label = classes_dict[class_name]cx = (x2+x)*0.5 / widthcy = (y2+y)*0.5 / heightw = (x2-x)*1. / widthh = (y2-y)*1. / heightline = "%s %.6f %.6f %.6f %.6f\n" % (label, cx, cy, w, h)lines.append(line)
正式得转化代码如下：
# 源数据地址：F:\datas\遥感数据集\NWPU VHR-10\oimport osimport numpy as npfrom PIL import Imagedef save_txt(save_folder, save_name, yolo_objects):save_path = os.path.join(save_folder, save_name)yolo_objects = np.array(yolo_objects)# todo numpy.savetxt（fname，X，fmt ='％。18e'，delimiter =''，# np.savetxt('data/task.txt', self.task, fmt="%d", delimiter=" ")# np.savetxt(save_path, yolo_objects, fmt ='％s', delimiter ='\n')np.savetxt(save_path, yolo_objects, fmt="%d %.18f %.18f %.18f %.18f", delimiter=' ')def parse_img(img_path):img = Image.open(img_path)# 通过size函数返回的是wHprint(img.size)np_img = np.array(img)# pillow 读取图片是HWCprint(np_img.shape)# 本地文件夹：E:\dddata\NWPU VHR-10 dataset# 解析标注文件def parse_txt(img_path, ann_path):objects = []ann_lines = np.loadtxt(ann_path, dtype=str, delimiter=',')n_dim = ann_lines.ndim# print(ann_lines.ndim)# print(ann_lines)img = Image.open(img_path)width, height = img.sizeif n_dim == 1:object = []x_min = int(ann_lines[0][1:])y_min = int(ann_lines[1][:-1])x_max = int(ann_lines[2][1:])y_max = int(ann_lines[3][:-1])label_id = int(ann_lines[4])object.append(x_min)object.append(y_min)object.append(x_max)object.append(y_max)object.append(label_id)object.append(width)object.append(height)objects.append(object)# print(x_min, y_min, x_max, y_max, lable_id)else:for ann_line in ann_lines:object = []x_min = int(ann_line[0][1:])y_min = int(ann_line[1][:-1])x_max = int(ann_line[2][1:])y_max = int(ann_line[3][:-1])label_id = int(ann_line[4])object.append(x_min)object.append(y_min)object.append(x_max)object.append(y_max)object.append(label_id)object.append(width)object.append(height)objects.append(object)# print(objects)return objects# 首先从原始文件中提取出文件路径, 文件名，以及图片大小等信息def get_origin_data(img_folder, ann_folder):# 通过标注文件去寻找原始图片文件，返回为字典的形式# key：图片路径 value: 目标信息数组imgpath_objects = []img_names = os.listdir(img_folder)for img_name in img_names:ann_name = img_name.split(".")[0] + '.txt'img_path = os.path.join(img_folder, img_name)ann_path = os.path.join(ann_folder, ann_name)# 读取txt文件objects = parse_txt(img_path, ann_path)# 构建字典one_imgpath_objects = {}one_imgpath_objects[img_path] = objectsimgpath_objects.append(one_imgpath_objects)# 返回的是字典数组，每个字典文件的key表示的图片路径，value是这个图片object的值return imgpath_objectsdef set_yolo_txt(imgpath_objects, save_folder = "E:/dddata/oher_format/yolo"):# 写成for循环，保存为yolo的形式，结束for one_imgpath_objects in imgpath_objects:# print(one_imgpath_objects)for img_path, objects in one_imgpath_objects.items():img_name = img_path.split(".")[0].split("\\")[-1]yolo_objects = []for object in objects:yolo_object = []x_min = object[0]y_min = object[1]x_max = object[2]y_max = object[3]label_id = object[4]width = object[5]height = object[6]center_x = (x_max+x_min)*0.5/widthcenter_y = (y_max+y_min)*0.5/heightw = (x_max-x_min)*1./widthh = (y_max-y_min)*1./heightyolo_object.append(label_id)yolo_object.append(center_x)yolo_object.append(center_y)yolo_object.append(w)yolo_object.append(h)yolo_objects.append(yolo_object)print(yolo_objects)save_name = img_name + '.txt'# save_folder = "F:/datas/遥感数据集/NWPU VHR-10/yolo"save_txt(save_folder, save_name, yolo_objects)if __name__ == '__main__':# 1. 返回字典信息imgpath_objects = get_origin_data(img_folder="E:/dddata/NWPU VHR-10 dataset/positive image set", ann_folder="E:/dddata/NWPU VHR-10 dataset/ground truth")# 2. 对字典信息进行处理,保存再txt文件中set_yolo_txt(imgpath_objects = imgpath_objects, save_folder="E:/dddata/oher_format/yolo")# parse_txt('F:/datas/遥感数据集/NWPU VHR-10/o/ground truth/003.txt')# parse_img("003.jpg")# parse_img("test.jpg")
VOC格式
voc形式的数据集一般包含以下几个文件夹
AnnotationsImageSetsJPEGImagesSegmentationClassSegmentationObject
对于目标检测的任务而言，一般只需要用到Annotations文件夹和JPEGImages文件夹，其中测试图片和训练图片都保存在一个目录下面，通过ImageSets文件夹中Main子文件夹下面的txt文件来对测试图片和验证图片进行解析。
JPEGImages主要提供的是PASCAL VOC所提供的所有的图片信息，包括训练图片，测试图片
这些图像就是用来进行训练和测试验证的图像数据。
Annotations主要存放xml格式的标签文件，每个xml对应JPEGImage中的一张图片
ImageSetsMain图像物体识别的数据，总共20类, 需要保证train val没有交集
SegmentationObject & SegmentationClass保存的是物体分割后的数据，在物体识别中没有用到
下面是标注文件的示例：
<annotation> <folder>VOC</folder> <filename>_000392.jpg</filename> //文件名 <source> //图像来源（不重要） <database>The VOC Database</database> <annotation>PASCAL VOC</annotation> <image>flickr</image> </source> <size> //图像尺寸（长宽以及通道数） <width>500</width> <height>332</height> <depth>3</depth> </size> <segmented>1</segmented> //是否用于分割（在图像物体识别中01无所谓） <object>//检测到的物体 <name>horse</name> //物体类别 <pose>Right</pose> //拍摄角度 <truncated>0</truncated>//是否被截断（0表示完整） <difficult>0</difficult>//目标是否难以识别（0表示容易识别） <bndbox> //bounding-box（包含左下角和右上角xy坐标） <xmin>100</xmin> <ymin>96</ymin> <xmax>355</xmax> <ymax>324</ymax> </bndbox> </object> <object> //检测到多个物体 <name>person</name> <pose>Unspecified</pose> <truncated>0</truncated> <difficult>0</difficult> <bndbox> <xmin>198</xmin> <ymin>58</ymin> <xmax>286</xmax> <ymax>197</ymax> </bndbox> </object> </annotation>
一般而言，有图片的名称、长和宽等信息，之后一张图片对应有多个object，每个object中包含了类名、然后三个对于目标检测而言不重要的信息，以及bndbox信息，bndbox信息十分重要，主要是左上角的坐标和右下角的坐标。
通过下列代码讲原始的标注信息转化为xml形式：
import osimport numpy as npfrom PIL import Imageimport xml.etree.ElementTree as ETfrom xml.dom.minidom import *# from xml.dom.minidom import# voc格式的目标需要转化为xml文件# todo 数据集类名classes_names = ['airplane', 'ship', 'storage tank', 'baseball diamond', 'tennis court', 'basketball court', 'ground track field', 'harbor', 'bridge', 'vehicle']# 1. 前面保持不变，还是把标注的基本信息转化回来# 本地文件夹：E:\dddata\NWPU VHR-10 dataset# 解析标注文件def parse_txt(img_path, ann_path):objects = []ann_lines = np.loadtxt(ann_path, dtype=str, delimiter=',')n_dim = ann_lines.ndim# print(ann_lines.ndim)# print(ann_lines)img = Image.open(img_path)width, height = img.sizeif n_dim == 1:object = []x_min = int(ann_lines[0][1:])y_min = int(ann_lines[1][:-1])x_max = int(ann_lines[2][1:])y_max = int(ann_lines[3][:-1])label_id = int(ann_lines[4])object.append(x_min)object.append(y_min)object.append(x_max)object.append(y_max)object.append(label_id)object.append(width)object.append(height)objects.append(object)# print(x_min, y_min, x_max, y_max, lable_id)else:for ann_line in ann_lines:object = []x_min = int(ann_line[0][1:])y_min = int(ann_line[1][:-1])x_max = int(ann_line[2][1:])y_max = int(ann_line[3][:-1])label_id = int(ann_line[4])object.append(x_min)object.append(y_min)object.append(x_max)object.append(y_max)object.append(label_id)object.append(width)object.append(height)objects.append(object)# print(objects)return objects# 首先从原始文件中提取出文件路径, 文件名，以及图片大小等信息def get_origin_data(img_folder, ann_folder):# 通过标注文件去寻找原始图片文件，返回为字典的形式# key：图片路径 value: 目标信息数组imgpath_objects = []img_names = os.listdir(img_folder)for img_name in img_names:ann_name = img_name.split(".")[0] + '.txt'img_path = os.path.join(img_folder, img_name)ann_path = os.path.join(ann_folder, ann_name)# 读取txt文件objects = parse_txt(img_path, ann_path)# 构建字典one_imgpath_objects = {}one_imgpath_objects[img_path] = objectsimgpath_objects.append(one_imgpath_objects)# 返回的是字典数组，每个字典文件的key表示的图片路径，value是这个图片object的值return imgpath_objects# 美化xml文件def pretty_xml(element, indent, newline, level=0): # elemnt为传进来的Elment类，参数indent用于缩进，newline用于换行if element: # 判断element是否有子元素if (element.text is None) or element.text.isspace(): # 如果element的text没有内容element.text = newline + indent * (level + 1)else:element.text = newline + indent * (level + 1) + element.text.strip() + newline + indent * (level + 1)# else: # 此处两行如果把注释去掉，Element的text也会另起一行# element.text = newline + indent * (level + 1) + element.text.strip() + newline + indent * leveltemp = list(element) # 将element转成listfor subelement in temp:if temp.index(subelement) < (len(temp) - 1): # 如果不是list的最后一个元素，说明下一个行是同级别元素的起始，缩进应一致subelement.tail = newline + indent * (level + 1)else: # 如果是list的最后一个元素， 说明下一行是母元素的结束，缩进应该少一个subelement.tail = newline + indent * levelpretty_xml(subelement, indent, newline, level=level + 1) # 对子元素进行递归操作# 写入xml文件def write_xml(img_name, width, height, object_dicts, save_path, folder='NWPU VHR-10 dataset'):'''object_dict = {'name': classes[int(object_category)],'truncated': int(truncation),'difficult': int(occlusion),'xmin':int(bbox_left),'ymin':int(bbox_top),'xmax':int(bbox_left) + int(bbox_width),'ymax':int(bbox_top) + int(bbox_height)}'''doc = Documentroot = ET.Element('Annotation')ET.SubElement(root, 'folder').text = folderET.SubElement(root, 'filename').text = img_namesize_node = ET.SubElement(root, 'size')ET.SubElement(size_node, 'width').text = str(width)ET.SubElement(size_node, 'height').text = str(height)ET.SubElement(size_node, 'depth').text = '3'for object_dict in object_dicts:object_node = ET.SubElement(root, 'object')ET.SubElement(object_node, 'name').text = object_dict['name']ET.SubElement(object_node, 'pose').text = 'Unspecified'ET.SubElement(object_node, 'truncated').text = str(object_dict['truncated'])ET.SubElement(object_node, 'difficult').text = str(object_dict['difficult'])bndbox_node = ET.SubElement(object_node, 'bndbox')ET.SubElement(bndbox_node, 'xmin').text = str(object_dict['xmin'])ET.SubElement(bndbox_node, 'ymin').text = str(object_dict['ymin'])ET.SubElement(bndbox_node, 'xmax').text = str(object_dict['xmax'])ET.SubElement(bndbox_node, 'ymax').text = str(object_dict['ymax'])pretty_xml(root, '\t', '\n')tree = ET.ElementTree(root)tree.write(save_path, encoding='utf-8')# 关键在于数据怎么写，前面基本一致def set_voc_xml(imgpath_objects, save_folder = "E:/dddata/oher_format/voc"):# 写成for循环，保存为yolo的形式，结束for one_imgpath_objects in imgpath_objects:# print(one_imgpath_objects)# 转化为xml所适用的字典列表# object_dict = {'name': classes[int(object_category)],#'truncated': int(truncation),#'difficult': int(occlusion),#'xmin': int(bbox_left),#'ymin': int(bbox_top),#'xmax': int(bbox_left) + int(bbox_width),#'ymax': int(bbox_top) + int(bbox_height)#}for img_path, objects in one_imgpath_objects.items():img_name = img_path.split(".")[0].split("\\")[-1]voc_objects = []width = 0height = 0for object in objects:# 这边实际上不需要太多的处理x_min = object[0]y_min = object[1]x_max = object[2]y_max = object[3]label_id = int(object[4])width_one = object[5]height_one = object[6]width = width_oneheight = height_onevoc_object = {'name': classes_names[int(label_id -1)],'truncated': 0,'difficult': 0,'xmin': int(x_min),'ymin': int(y_min),'xmax': int(x_max),'ymax': int(y_max)}voc_objects.append(voc_object)save_name = img_name + '.xml'save_path = os.path.join(save_folder, save_name)print(save_name)write_xml(img_name, width, height, voc_objects, save_path, folder='NWPU VHR-10 dataset')if __name__ == '__main__':imgpath_objects = get_origin_data(img_folder="E:/dddata/NWPU VHR-10 dataset/positive image set",ann_folder="E:/dddata/NWPU VHR-10 dataset/ground truth")# 2. 对字典信息进行处理,保存再txt文件中set_voc_xml(imgpath_objects=imgpath_objects, save_folder="E:/dddata/oher_format/voc")
coco格式
参考：/qq_41375609/article/details/94737915
coco数据集为json文件，一般包含5个字段
infoimagesannotationslicensescategories
这次我拿到的数据集是关于红外图像的数据集，每个字段的含义如下
info字段
info字段包含了数据集的基本信息，包括数据集的来源，提供者之类的，内容如下：
info: {"year": int,# 年份"version": str,# 版本"description": str, # 数据集描述"contributor": str,# 提供者"url": str,# 下载地址"date_created": datetime}
示例如下：
"info":{"description":"This is stable 1.0 version of the MS COCO dataset.","url":"http:\/\/","version":"1.0","year":,"contributor":"Microsoft COCO group","date_created":"-01-27 09:11:52.357475"},
info字段在写程序的时候一般不会使用到
licenses字段
licenses字段表明了图片的版权信息之类的，一般的程序中也用不到
licenses字段的结构如下：
license{"id": int,"name": str,"url": str,}
licenses字段的示例如下：
{"url":"http:\/\/\/licenses\/by-nc-sa\/2.0\/","id":1,"name":"Attribution-NonCommercial-ShareAlike License"}
images字段
images字段是整个json文件中最重要的字段之一，包含了图片的基本信息，包括图片的名称，宽高。images目录由多个image构成数组，可以遍历，结构如下：
image{"id": int,# 图片的ID编号（每张图片ID是唯一的）"width": int,#宽"height": int,#高"file_name": str,# 图片名"license": int,"flickr_url": str,# flickr网路地址"coco_url": str,# 网路地址路径"date_captured": datetime # 数据获取日期}
示例如下：
{"license":3,"file_name":"COCO_val_000000391895.jpg","coco_url":"http:\/\/\/images\/391895","height":360,"width":640,"date_captured":"-11-14 11:18:45","flickr_url":"http:\/\/\/8186\/8119368305_4e622c8349_z.jpg","id":391895}
每一个image的实例是一个dict。其中有一个id字段，代表的是图片的id，每一张图片具有唯一的一个独特的id。
annotations字段
存储图片的标注信息，结构如下：
annotation{"id": int, # 对象ID，因为每一个图像有不止一个对象，所以要对每一个对象编号（每个对象的ID是唯一的）"image_id": int,# 对应的图片ID（与images中的ID对应）"category_id": int,# 类别ID（与categories中的ID对应）"segmentation": RLE or [polygon],# 对象的边界点（边界多边形，此时iscrowd=0）。#segmentation格式取决于这个实例是一个单个的对象（即iscrowd=0，将使用polygons格式）还是一组对象（即iscrowd=1，将使用RLE格式）"area": float,# 区域面积"bbox": [x,y,width,height], # 定位边框 [x,y,w,h]"iscrowd": 0 or 1 #见下}
其中注意这里的bbox格式为[x,y,width,height], # 定位边框 [x,y,w,h]，是没有进行归一化的xywh，所以在进行yolo的转化时需要进行归一化的处理，示例如下：
比如cocotrain或cocoval数据集中标注的目标(类别)位置在 Annotations 中以 (x, y, width, height) 来进行表示，x，y表示bbox左上角位置，width, height表示bbox的宽和高。而YOLO训练或者进行验证的时候读取的标注格式是以 (xmin, ymin, xmax, ymax) 来进行表示，xmin, ymin表示bbox左上角位置， xmax, ymax表示bbox右下角位置。
{"segmentation": [[510.66,423.01,511.72,420.03,510.45......]],"area": 702.1057499999998,"iscrowd": 0,"image_id": 289343,"bbox": [473.07,395.93,38.65,28.67],"category_id": 18,"id": 1768}
categories字段
这个字段主要是记录annotations字段中的类别信息，结构如下：
{"supercategory": str,# 主类别"id": int,# 类对应的id （0 默认为背景）"name": str # 子类别}
示例如下：
{"supercategory": "person","id": 1,"name": "person"},{"supercategory": "vehicle","id": 2,"name": "bicycle"}
附录
voc与voc之间的转化公式
# object_dict = {'name': class_names_detrac[int(content[0])],#'truncated': 0,#'difficult': 0,#'xmin': int(center_x - bbox_width / 2),#'ymin': int(center_y - bbox_height / 2),#'xmax': int(center_x + bbox_width / 2),#'ymax': int(center_y + bbox_height / 2)#}
代码如下，处理的比较乱，后面再调整
import jsonimport copyimport numpy as npimport osimport shutil# 修正框# 有个问题是数据集中的名称是按照。。执行的def correct_objects(objects, width, height):yolo_objects = []for object in objects:# 原始数据是xywhx_min = object[0]y_min = object[1]box_w = object[2]box_h = object[3]if x_min < 0 or y_min < 0 or box_w < 0 or box_h < 0:passelse:x_max = x_min + box_wy_max = y_min + box_hlabel_id = object[4]center_x = (x_max + x_min) * 0.5 / widthcenter_y = (y_max + y_min) * 0.5 / heightwx = (x_max - x_min) * 1. / widthhx = (y_max - y_min) * 1. / heightyolo_object = []yolo_object.append(label_id)yolo_object.append(center_x)yolo_object.append(center_y)yolo_object.append(wx)yolo_object.append(hx)if (center_x < 1) and (center_y <1) and (wx<1) and (hx <1):yolo_objects.append(yolo_object)return yolo_objectsdef save_txt(save_folder, save_name, yolo_objects):save_path = os.path.join(save_folder, save_name)yolo_objects = np.array(yolo_objects)# todo numpy.savetxt（fname，X，fmt ='％。18e'，delimiter =''，# np.savetxt('data/task.txt', self.task, fmt="%d", delimiter=" ")# np.savetxt(save_path, yolo_objects, fmt ='％s', delimiter ='\n')np.savetxt(save_path, yolo_objects, fmt="%d %.18f %.18f %.18f %.18f", delimiter=' ')with open("aauRainSnow-thermal.json", "r") as load_f:load_dict = json.load(load_f)# print(load_dict)print("字段信息：")for key, value in load_dict.items():print(key)# 首先把每个字段的信息记录并保存下来info = load_dict['info']images = load_dict['images']annotations = load_dict['annotations']licenses = load_dict['licenses']categories = load_dict['categories']# 这边的categories是从1开始的，应该转化为从0开始# print(categories)categories_list = []for category in categories:categories_list.append(category['name'])print("种类：{}种".format(len(categories_list)))# 首先对图片做一个遍历# 初始化image_path_objectstmp_dict_list = {}for image in images:file_name = image['file_name']cam_id = int(file_name.split('cam')[-1][0])if cam_id == 2:image_copy = copy.copy(image)image_copy['objects'] = []# 然后存在tmp数据中# 为了方便查询，把图片的id信息作为key进行保存tmp_dict = {}tmp_dict_list[image_copy['id']] = image_copy# tmp_dict_list.append(tmp_dict)# 遍历标注信息，把标注信息存入for ann in annotations:ann_img_id = ann['image_id']ann_bbox = ann['bbox']ann_category_id = ann['category_id'] - 1ann_bbox.append(ann_category_id)# 添加主信息中# print(tmp_dict_list[ann_img_id])tmp_dict_list[ann_img_id]['objects'].append(ann_bbox)# imgpath_objects# todo：在这里不用返回直接做写入操作：# imgpath_objects = []for value in tmp_dict_list.values():file_name = value["file_name"]file_just_name = file_name.split("/")[-1].split(".")[0]print(file_just_name)width = value['width']height = value['height']objects = value['objects']# 现在需要对objects做修正yolo_objects = correct_objects(objects, width, height)save_txt_name = file_just_name + '.txt'# print(yolo_objects)if len(yolo_objects) > 0:# todo 保存文本文件save_txt(save_folder="F:/datas/红外/AAU-RainSnow精简/aaurainsnow/v2/yolo_anns", save_name=save_txt_name, yolo_objects=np.array(yolo_objects))# todo 移动图片文件img_name = file_name.split("/")[-1]image_target_path = os.path.join("F:/datas/红外/AAU-RainSnow精简/aaurainsnow/v2/yolo_images", img_name)image_src_path = os.path.join("F:/datas/红外/AAU-RainSnow精简/aaurainsnow", file_name)shutil.copy(image_src_path, image_target_path)
致谢
/p/255256266?utm_source=wechat_session
xxx
