• pycocotools库的使用

    1.组成模块

    pycocotools下有三个模块：coco、cocoeval、mask、_mask。

    （1）coco模块： 

    1. # The following API functions are defined:
    2. #  COCO       - COCO api class that loads COCO annotation file and prepare data structures.
    3. #  getAnnIds  - Get ann ids that satisfy given filter conditions.
    4. #  getCatIds  - Get cat ids that satisfy given filter conditions.
    5. #  getImgIds  - Get img ids that satisfy given filter conditions.
    6. #  loadAnns   - Load anns with the specified ids.
    7. #  loadCats   - Load cats with the specified ids.
    8. #  loadImgs   - Load imgs with the specified ids.
    9. #  annToMask  - Convert segmentation in an annotation to binary mask.
    10. #  showAnns   - Display the specified annotations.
    11. #  loadRes    - Load algorithm results and create API for accessing them.
    12. #  download   - Download COCO images from mscoco.org server.
    13. # Throughout the API "ann"=annotation, "cat"=category, and "img"=image.
    14. # Help on each functions can be accessed by: "help COCO>function".

    COCO类定义了10个方法：

    （1）获取标注id：

    1. def getAnnIds(self, imgIds=[], catIds=[], areaRng=[], iscrowd=None):
    2.         """
    3.         Get ann ids that satisfy given filter conditions. default skips that filter
    4.         :param imgIds  (int array)     : get anns for given imgs
    5.                catIds  (int array)     : get anns for given cats
    6.                areaRng (float array)   : get anns for given area range (e.g. [0 inf])
    7.                iscrowd (boolean)       : get anns for given crowd label (False or True)
    8.         :return: ids (int array)       : integer array of ann ids
    9.         """

    （2）获取类别id：

    1. def getCatIds(self, catNms=[], supNms=[], catIds=[]):
    2.         """
    3.         filtering parameters. default skips that filter.
    4.         :param catNms (str array)  : get cats for given cat names
    5.         :param supNms (str array)  : get cats for given supercategory names
    6.         :param catIds (int array)  : get cats for given cat ids
    7.         :return: ids (int array)   : integer array of cat ids
    8.         """

    （3）获取图片id：

    1. def getImgIds(self, imgIds=[], catIds=[]):
    2.         '''
    3.         Get img ids that satisfy given filter conditions.
    4.         :param imgIds (int array) : get imgs for given ids
    5.         :param catIds (int array) : get imgs with all given cats
    6.         :return: ids (int array)  : integer array of img ids
    7.         '''

    （4）加载标注信息：

    1. def loadAnns(self, ids=[]):
    2.         """
    3.         Load anns with the specified ids.
    4.         :param ids (int array)       : integer ids specifying anns
    5.         :return: anns (object array) : loaded ann objects
    6.         """

    （5）加载类别：

    1. def loadCats(self, ids=[]):
    2.         """
    3.         Load cats with the specified ids.
    4.         :param ids (int array)       : integer ids specifying cats
    5.         :return: cats (object array) : loaded cat objects
    6.         """

    （6）加载图片：

    1. def loadImgs(self, ids=[]):
    2.         """
    3.         Load anns with the specified ids.
    4.         :param ids (int array)       : integer ids specifying img
    5.         :return: imgs (object array) : loaded img objects
    6.         """

    （7）用matplotlib在图片上显示标注：

    1. def showAnns(self, anns):
    2.         """
    3.         Display the specified annotations.
    4.         :param anns (array of object): annotations to display
    5.         :return: None
    6.         """

    （8）加载结果文件：

    1. def loadRes(self, resFile):
    2.         """
    3.         Load result file and return a result api object.
    4.         :param   resFile (str)     : file name of result file
    5.         :return: res (obj)         : result api object
    6.         """

    （9）下载数据集：

    1. def download(self, tarDir = None, imgIds = [] ):
    2.         '''
    3.         Download COCO images from mscoco.org server.
    4.         :param tarDir (str): COCO results directory name
    5.                imgIds (list): images to be downloaded
    6.         :return:
    7.         '''

    （10）ann(polygons, uncompressed RLE)转为rle格式(0表示背景，1表示分割区域)：

    1. def annToRLE(self, ann):
    2.         """
    3.         Convert annotation which can be polygons, uncompressed RLE to RLE.
    4.         :return: binary mask (numpy 2D array)
    5.         """

    （11）polygons, uncompressed RLE, or RLE 转mask：

    1. def annToMask(self, ann):
    2.         """
    3.         Convert annotation which can be polygons, uncompressed RLE, or RLE to binary mask.
    4.         :return: binary mask (numpy 2D array)
    5.         """

     

    2、mask模块下定义了四个函数：

    1. def encode(bimask)：
    2. def decode(rleObjs):
    3. def area(rleObjs):
    4. def toBbox(rleObjs):

    3、cocoeval模块定义了COCOeval和Params类：

    1.     # The usage for CocoEval is as follows:
    2.     #  cocoGt=..., cocoDt=...       # load dataset and results
    3.     #  E = CocoEval(cocoGt,cocoDt); # initialize CocoEval object
    4.     #  E.params.recThrs = ...;      # set parameters as desired
    5.     #  E.evaluate();                # run per image evaluation
    6.     #  E.accumulate();              # accumulate per image results
    7.     #  E.summarize();               # display summary metrics of results

    2.使用pycocotools加载coco数据集

     1.实例化coco类

    从coco类的源码中，我们可以看到，初始化方法中执行了createIndex方法，其中返回

    字典anns：以标注id为keys，标注信息为values的字典

    字典imgs：以图片id为健，图片信息为值的字典

    字典imgToAnns：以图片id为健，标注信息为值（列表）的字典

    字典cats：以类别id为健，类别信息为值的字典

    字典catToImgs：以种类id为健，图片id（list）的字典

    1. class COCO:
    2.     def __init__(self, annotation_file=None):
    3.         """
    4.         Constructor of Microsoft COCO helper class for reading and visualizing annotations.
    5.         :param annotation_file (str): location of annotation file
    6.         :param image_folder (str): location to the folder that hosts images.
    7.         :return:
    8.         """
    9.         # load dataset
    10.         self.dataset,self.anns,self.cats,self.imgs = dict(),dict(),dict(),dict()
    11.         self.imgToAnns, self.catToImgs = defaultdict(list), defaultdict(list)
    12.         if not annotation_file == None:
    13.             print('loading annotations into memory...')
    14.             tic = time.time()
    15.             with open(annotation_file, 'r') as f:
    16.                 dataset = json.load(f)
    17.             assert type(dataset)==dict, 'annotation file format {} not supported'.format(type(dataset))
    18.             print('Done (t={:0.2f}s)'.format(time.time()- tic))
    19.             self.dataset = dataset
    20.             self.createIndex()
    21.  
    22.     def createIndex(self):
    23.         # create index
    24.         print('creating index...')
    25.         anns, cats, imgs = {}, {}, {}
    26.         imgToAnns,catToImgs = defaultdict(list),defaultdict(list)
    27.         if 'annotations' in self.dataset:
    28.             for ann in self.dataset['annotations']:
    29.                 imgToAnns[ann['image_id']].append(ann)
    30.                 anns[ann['id']] = ann
    31.  
    32.         if 'images' in self.dataset:
    33.             for img in self.dataset['images']:
    34.                 imgs[img['id']] = img
    35.  
    36.         if 'categories' in self.dataset:
    37.             for cat in self.dataset['categories']:
    38.                 cats[cat['id']] = cat
    39.  
    40.         if 'annotations' in self.dataset and 'categories' in self.dataset:
    41.             for ann in self.dataset['annotations']:
    42.                 catToImgs[ann['category_id']].append(ann['image_id'])
    43.  
    44.         print('index created!')
    45.  
    46.         # create class members
    47.         self.anns = anns
    48.         self.imgToAnns = imgToAnns
    49.         self.catToImgs = catToImgs
    50.         self.imgs = imgs
    51.         self.cats = cats

     2.读取数据

    我们以tensorflow读取数据为例，pytorch也类似，tensorflow重写的方法keras.utils.Sequence与pytorch需要重写的方法dataloader类似

    1. class COCODetection(Sequence):
    2.     def __init__(self, image_path, coco, num_classes, anchors, batch_size, config, COCO_LABEL_MAP={}, augmentation=None):
    3.         self.image_path     = image_path
    4.  
    5.         self.coco           = coco
    6.         self.ids            = list(self.coco.imgToAnns.keys())
    7.  
    8.         self.num_classes    = num_classes
    9.         self.anchors        = anchors
    10.         self.batch_size     = batch_size
    11.         self.config         = config
    12.  
    13.         self.augmentation   = augmentation
    14.  
    15.         self.label_map      = COCO_LABEL_MAP
    16.         self.length         = len(self.ids)
    17.  
    18.     def __getitem__(self, index):
    19.         for i, global_index in enumerate(range(index * self.batch_size, (index + 1) * self.batch_size)):  
    20.             global_index = global_index % self.length
    21.  
    22.             image, boxes, mask_gt, num_crowds, image_id = self.pull_item(global_index)
    23.             #------------------------------#
    24.             #   获得种类
    25.             #------------------------------#
    26.             class_ids   = boxes[:,  -1]
    27.             #------------------------------#
    28.             #   获得框的坐标
    29.             #------------------------------#
    30.             boxes       = boxes[:, :-1]
    31.  
    32.             image, image_meta, gt_class_ids, gt_boxes, gt_masks = \
    33.             load_image_gt(image, mask_gt, boxes, class_ids, image_id, self.config, use_mini_mask=self.config.USE_MINI_MASK)
    34.  
    35.             #------------------------------#
    36.             #   初始化用于训练的内容
    37.             #------------------------------#
    38.             if i == 0:
    39.                 batch_image_meta    = np.zeros((self.batch_size,) + image_meta.shape, dtype=image_meta.dtype)
    40.                 batch_rpn_match     = np.zeros([self.batch_size, self.anchors.shape[0], 1], dtype=np.int32)
    41.                 batch_rpn_bbox      = np.zeros([self.batch_size, self.config.RPN_TRAIN_ANCHORS_PER_IMAGE, 4], dtype=np.float32)
    42.                 batch_images        = np.zeros((self.batch_size,) + image.shape, dtype=np.float32)
    43.                 batch_gt_class_ids  = np.zeros((self.batch_size, self.config.MAX_GT_INSTANCES), dtype=np.int32)
    44.                 batch_gt_boxes      = np.zeros((self.batch_size, self.config.MAX_GT_INSTANCES, 4), dtype=np.int32)
    45.                 batch_gt_masks      = np.zeros((self.batch_size, gt_masks.shape[0], gt_masks.shape[1], self.config.MAX_GT_INSTANCES), dtype=gt_masks.dtype)
    46.  
    47.             if not np.any(gt_class_ids > 0):
    48.                 continue
    49.         
    50.             # RPN Targets
    51.             rpn_match, rpn_bbox = build_rpn_targets(image.shape, self.anchors, gt_class_ids, gt_boxes, self.config)
    52.         
    53.             #-----------------------------------------------------------------------#
    54.             #   如果某张图片里面物体的数量大于最大值的话，则进行筛选，防止过大
    55.             #-----------------------------------------------------------------------#
    56.             if gt_boxes.shape[0] > self.config.MAX_GT_INSTANCES:
    57.                 ids = np.random.choice(
    58.                     np.arange(gt_boxes.shape[0]), self.config.MAX_GT_INSTANCES, replace=False)
    59.                 gt_class_ids = gt_class_ids[ids]
    60.                 gt_boxes = gt_boxes[ids]
    61.                 gt_masks = gt_masks[:, :, ids]
    62.  
    63.             #------------------------------#
    64.             #   将当前信息加载进batch
    65.             #------------------------------#
    66.             batch_image_meta[i]                             = image_meta
    67.             batch_rpn_match[i]                              = rpn_match[:, np.newaxis]
    68.             batch_rpn_bbox[i]                               = rpn_bbox
    69.             batch_images[i]                                 = preprocess_input(image.astype(np.float32))
    70.             batch_gt_class_ids[i, :gt_class_ids.shape[0]]   = gt_class_ids
    71.             batch_gt_boxes[i, :gt_boxes.shape[0]]           = gt_boxes
    72.             batch_gt_masks[i, :, :, :gt_masks.shape[-1]]    = gt_masks
    73.         return [batch_images, batch_image_meta, batch_rpn_match, batch_rpn_bbox, batch_gt_class_ids, batch_gt_boxes, batch_gt_masks], \
    74.                [np.zeros(self.batch_size), np.zeros(self.batch_size), np.zeros(self.batch_size), np.zeros(self.batch_size), np.zeros(self.batch_size)]
    75.  
    76.     def __len__(self):
    77.         return math.ceil(len(self.ids) / float(self.batch_size))
    78.  
    79.     def pull_item(self, index):
    80.         #------------------------------#
    81.         #   载入coco序号
    82.         #   根据coco序号载入目标信息
    83.         #------------------------------#
    84.         image_id    = self.ids[index]
    85.         target      = self.coco.loadAnns(self.coco.getAnnIds(imgIds = image_id))
    86.  
    87.         #------------------------------#
    88.         #   根据目标信息判断是否为
    89.         #   iscrowd
    90.         #------------------------------#
    91.         target      = [x for x in target if not ('iscrowd' in x and x['iscrowd'])]
    92.         crowd       = [x for x in target if ('iscrowd' in x and x['iscrowd'])]
    93.         num_crowds  = len(crowd)
    94.         #------------------------------#
    95.         #   将不是iscrowd的目标
    96.         #       是iscrowd的目标进行堆叠
    97.         #------------------------------#
    98.         target      += crowd
    99.  
    100.         image_path  = osp.join(self.image_path, self.coco.loadImgs(image_id)[0]['file_name'])
    101.         image       = Image.open(image_path)
    102.         image       = cvtColor(image)
    103.         image       = np.array(image, np.float32)
    104.         height, width, _ = image.shape
    105.  
    106.         if len(target) > 0:
    107.             masks = np.array([self.coco.annToMask(obj).reshape(-1) for obj in target], np.float32)
    108.             masks = masks.reshape((-1, height, width)) 
    109.  
    110.             boxes_classes = []
    111.             for obj in target:
    112.                 bbox        = obj['bbox']
    113.                 final_box   = [bbox[0], bbox[1], bbox[0] + bbox[2], bbox[1] + bbox[3], self.label_map[obj['category_id']]]
    114.                 boxes_classes.append(final_box)
    115.             boxes_classes = np.array(boxes_classes, np.float32)
    116.             boxes_classes[:, [0, 2]] /= width
    117.             boxes_classes[:, [1, 3]] /= height
    118.  
    119.         if self.augmentation is not None:
    120.             if len(boxes_classes) > 0:
    121.                 image, masks, boxes, labels = self.augmentation(image, masks, boxes_classes[:, :4], {'num_crowds': num_crowds, 'labels': boxes_classes[:, 4]})
    122.                 num_crowds  = labels['num_crowds']
    123.                 labels      = labels['labels']
    124.                 if num_crowds > 0:
    125.                     labels[-num_crowds:] = -1
    126.                 boxes       = np.concatenate([boxes, np.expand_dims(labels, axis=1)], -1)
    127.         
    128.         masks               = np.transpose(masks, [1, 2, 0])
    129.         outboxes            = np.zeros_like(boxes)
    130.         outboxes[:, [0, 2]] = boxes[:, [1, 3]] * self.config.IMAGE_SHAPE[0]
    131.         outboxes[:, [1, 3]] = boxes[:, [0, 2]] * self.config.IMAGE_SHAPE[1]
    132.         outboxes[:, -1]     = boxes[:, -1]
    133.         outboxes            = np.array(outboxes, np.int)
    134.         return image, outboxes, masks, num_crowds, image_id

     

  • 相关阅读:
    再理解springboot那些注册与回调、监控与统计等命名规范，就可以读懂70%的springboot源代码
    让历史文化“活”起来，北京河图“万象中轴”助力打造北京城市金名片
    Java_自定义实体类的列表List＜T＞调用remove()失败讲解
    Denpendcy Injection 8.0新功能——KeyedService
    【微信小程序】自定义组件(一)
    基于Rspack实现大仓应用构建提效实践｜得物技术
    SQL: MAX Function
    【AI工程】08-MLOps工具-在Charmed Kubeflow上运行MindSpore
    免费调用快递鸟物流跟踪轨迹订阅接口技术文档
    ubuntu22.04.1 新装后的常用设置
  • 原文地址：https://blog.csdn.net/qq_52053775/article/details/126197172