def __init__(self,
dataset_dir,
subset,
flip_ratio=0,
pad_mode='fixed',
mean=(0, 0, 0),
std=(1, 1, 1),
scale=(1024, 800),
debug=False):
"""Load a subset of the COCO dataset.
Attributes
---
dataset_dir: The root directory of the COCO dataset.
subset: What to load (train, val).
flip_ratio: Float. The ratio of flipping an image and its bounding boxes.
pad_mode: Which padded method to use (fixed, non-fixed)
mean: Tuple. Image mean.
std: Tuple. Image standard deviation.
scale: Tuple of two integers.
"""
if subset not in ['train', 'val']:
raise AssertionError('subset must be "train" or "val".')
self.coco = COCO("{}/annotations/instances_{}2017.json".format(
dataset_dir, subset))
# get the mapping from original category ids to labels
self.cat_ids = self.coco.getCatIds()
self.cat2label = {
cat_id: i + 1
for i, cat_id in enumerate(self.cat_ids)
}
self.label2cat = {
i + 1: cat_id
for i, cat_id in enumerate(self.cat_ids)
}
self.img_ids, self.img_infos = self._filter_imgs()
if debug:
self.img_ids, self.img_infos = self.img_ids[:
20], self.img_infos[:
20]
self.image_dir = "{}/{}2017".format(dataset_dir, subset)
self.flip_ratio = flip_ratio
if pad_mode in ['fixed', 'non-fixed']:
self.pad_mode = pad_mode
elif subset == 'train':
self.pad_mode = 'fixed'
else:
self.pad_mode = 'non-fixed'
self.img_transform = transforms.ImageTransform(scale, mean, std,
pad_mode)
self.bbox_transform = transforms.BboxTransform()
def __init__(self,
dataset_dir,
subset,
flip_ratio=0,
pad_mode='fixed',
config=None,
debug=False):
'''Load a subset of the COCO dataset.
Attributes
---
dataset_dir: The root directory of the COCO dataset.
subset: What to load (train, val).
flip_ratio: Float. The ratio of flipping an image and its bounding boxes.
pad_mode: Which padded method to use (fixed, non-fixed)
scale: Tuple of two integers.
'''
self.config = config
self.debug = debug
if subset not in ['train', 'val']:
raise AssertionError('subset must be "train" or "val".')
self.coco = COCO(
f"{dataset_dir}/annotations/instances_{subset}2017.json")
# get the mapping from original category ids to labels
self.cat_ids = self.coco.getCatIds()
self.cat2label = {
cat_id: i + 1
for i, cat_id in enumerate(self.cat_ids)
}
self.img_ids, self.img_infos = self._filter_imgs()
if debug:
self.img_ids, self.img_infos = self.img_ids[:
50], self.img_infos[:
50]
self.image_dir = f"{dataset_dir}/{subset}2017"
self.flip_ratio = flip_ratio
if pad_mode in ['fixed', 'non-fixed']:
self.pad_mode = pad_mode
elif subset == 'train':
self.pad_mode = 'fixed'
else:
self.pad_mode = 'non-fixed'
self.img_transform = transforms.ImageTransform(
(config.IMAGE_MIN_DIM, config.IMAGE_MAX_DIM), config.MEAN_PIXEL,
config.STD_PIXEL, pad_mode)
self.bbox_transform = transforms.BboxTransform()
def __init__(self, dataset_dir, subset,
flip_ratio=0,
pad_mode='fixed',
mean=(0, 0, 0),
std=(1, 1, 1),
scale=(1024, 800),
debug=False):
"""加载coco数据集
Attributes
---
dataset_dir: 数据集根目录
subset: 'train'或者'val'
flip_ratio: Float. 按照flip_ratio为概率决定图片是否进行左右翻转
pad_mode: 使用那种pad模式(fixed, non-fixed)，如果是fixed则会填充至scale中的较大值，
否则以高、宽都达到64的倍数为目的进行最小填充
mean: [3] 均值，用于归一化
std: [3] 标准差，用于归一化
scale: [2] 用于决定放缩大小
"""
if subset not in ['train', 'val', 'test']:
raise AssertionError('subset must be "train" or "val" or "test".')
self.coco = COCO("{}/{}/{}.json".format(dataset_dir, subset, subset))
self.image_dir = "{}/{}/images".format(dataset_dir, subset) # 图片路径
self.flip_ratio = flip_ratio
if pad_mode in ['fixed', 'non-fixed']:
self.pad_mode = pad_mode
elif subset == 'train': # 对于训练集，进行固定填充，即将图片宽、高都填充至固定值（正方形）
self.pad_mode = 'fixed'
else: # 对于非训练集，进行非固定填充，即讲图片宽、高都填充至一个固定值的整数倍
self.pad_mode = 'non-fixed'
self.cat_ids = self.coco.getCatIds() # 标注文件中的所有类别号
self.cat2label = { # 标注文件中的类别编号：新的类别编号
cat_id: i + 1
for i, cat_id in enumerate(self.cat_ids)
}
self.label2cat = { # 新的类别编号：标注文件中的类别编号
i + 1: cat_id
for i, cat_id in enumerate(self.cat_ids)
}
self.img_ids, self.img_infos = self._filter_imgs() # 图片编号以及相应的标签信息
if debug:
self.img_ids, self.img_infos = self.img_ids[:50], self.img_infos[:50]
self.img_transform = transforms.ImageTransform(scale, mean, std, pad_mode) # 图片预处理类
self.bbox_transform = transforms.BboxTransform() # 边界框相应处理类