det_transforms.py 75 KB

1234567891011121314151617181920212223242526272829303132333435363738394041424344454647484950515253545556575859606162636465666768697071727374757677787980818283848586878889909192939495969798991001011021031041051061071081091101111121131141151161171181191201211221231241251261271281291301311321331341351361371381391401411421431441451461471481491501511521531541551561571581591601611621631641651661671681691701711721731741751761771781791801811821831841851861871881891901911921931941951961971981992002012022032042052062072082092102112122132142152162172182192202212222232242252262272282292302312322332342352362372382392402412422432442452462472482492502512522532542552562572582592602612622632642652662672682692702712722732742752762772782792802812822832842852862872882892902912922932942952962972982993003013023033043053063073083093103113123133143153163173183193203213223233243253263273283293303313323333343353363373383393403413423433443453463473483493503513523533543553563573583593603613623633643653663673683693703713723733743753763773783793803813823833843853863873883893903913923933943953963973983994004014024034044054064074084094104114124134144154164174184194204214224234244254264274284294304314324334344354364374384394404414424434444454464474484494504514524534544554564574584594604614624634644654664674684694704714724734744754764774784794804814824834844854864874884894904914924934944954964974984995005015025035045055065075085095105115125135145155165175185195205215225235245255265275285295305315325335345355365375385395405415425435445455465475485495505515525535545555565575585595605615625635645655665675685695705715725735745755765775785795805815825835845855865875885895905915925935945955965975985996006016026036046056066076086096106116126136146156166176186196206216226236246256266276286296306316326336346356366376386396406416426436446456466476486496506516526536546556566576586596606616626636646656666676686696706716726736746756766776786796806816826836846856866876886896906916926936946956966976986997007017027037047057067077087097107117127137147157167177187197207217227237247257267277287297307317327337347357367377387397407417427437447457467477487497507517527537547557567577587597607617627637647657667677687697707717727737747757767777787797807817827837847857867877887897907917927937947957967977987998008018028038048058068078088098108118128138148158168178188198208218228238248258268278288298308318328338348358368378388398408418428438448458468478488498508518528538548558568578588598608618628638648658668678688698708718728738748758768778788798808818828838848858868878888898908918928938948958968978988999009019029039049059069079089099109119129139149159169179189199209219229239249259269279289299309319329339349359369379389399409419429439449459469479489499509519529539549559569579589599609619629639649659669679689699709719729739749759769779789799809819829839849859869879889899909919929939949959969979989991000100110021003100410051006100710081009101010111012101310141015101610171018101910201021102210231024102510261027102810291030103110321033103410351036103710381039104010411042104310441045104610471048104910501051105210531054105510561057105810591060106110621063106410651066106710681069107010711072107310741075107610771078107910801081108210831084108510861087108810891090109110921093109410951096109710981099110011011102110311041105110611071108110911101111111211131114111511161117111811191120112111221123112411251126112711281129113011311132113311341135113611371138113911401141114211431144114511461147114811491150115111521153115411551156115711581159116011611162116311641165116611671168116911701171117211731174117511761177117811791180118111821183118411851186118711881189119011911192119311941195119611971198119912001201120212031204120512061207120812091210121112121213121412151216121712181219122012211222122312241225122612271228122912301231123212331234123512361237123812391240124112421243124412451246124712481249125012511252125312541255125612571258125912601261126212631264126512661267126812691270127112721273127412751276127712781279128012811282128312841285128612871288128912901291129212931294129512961297129812991300130113021303130413051306130713081309131013111312131313141315131613171318131913201321132213231324132513261327132813291330133113321333133413351336133713381339134013411342134313441345134613471348134913501351135213531354135513561357135813591360136113621363136413651366136713681369137013711372137313741375137613771378137913801381138213831384138513861387138813891390139113921393139413951396139713981399140014011402140314041405140614071408140914101411141214131414141514161417141814191420142114221423142414251426142714281429143014311432143314341435143614371438143914401441144214431444144514461447144814491450145114521453145414551456145714581459146014611462146314641465146614671468146914701471147214731474147514761477147814791480148114821483148414851486148714881489149014911492149314941495149614971498149915001501150215031504150515061507150815091510151115121513151415151516151715181519152015211522152315241525152615271528152915301531153215331534153515361537153815391540154115421543154415451546154715481549155015511552155315541555155615571558155915601561156215631564156515661567156815691570157115721573157415751576157715781579158015811582158315841585158615871588158915901591159215931594159515961597159815991600160116021603160416051606160716081609161016111612161316141615161616171618161916201621162216231624162516261627162816291630163116321633163416351636163716381639
  1. # copyright (c) 2020 PaddlePaddle Authors. All Rights Reserved.
  2. #
  3. # Licensed under the Apache License, Version 2.0 (the "License");
  4. # you may not use this file except in compliance with the License.
  5. # You may obtain a copy of the License at
  6. #
  7. # http://www.apache.org/licenses/LICENSE-2.0
  8. #
  9. # Unless required by applicable law or agreed to in writing, software
  10. # distributed under the License is distributed on an "AS IS" BASIS,
  11. # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
  12. # See the License for the specific language governing permissions and
  13. # limitations under the License.
  14. try:
  15. from collections.abc import Sequence
  16. except Exception:
  17. from collections import Sequence
  18. import random
  19. import os.path as osp
  20. import numpy as np
  21. import cv2
  22. from PIL import Image, ImageEnhance
  23. from .imgaug_support import execute_imgaug
  24. from .ops import *
  25. from .box_utils import *
  26. import paddlex.utils.logging as logging
  27. class DetTransform:
  28. """检测数据处理基类
  29. """
  30. def __init__(self):
  31. pass
  32. class Compose(DetTransform):
  33. """根据数据预处理/增强列表对输入数据进行操作。
  34. 所有操作的输入图像流形状均是[H, W, C],其中H为图像高,W为图像宽,C为图像通道数。
  35. Args:
  36. transforms (list): 数据预处理/增强列表。
  37. Raises:
  38. TypeError: 形参数据类型不满足需求。
  39. ValueError: 数据长度不匹配。
  40. """
  41. def __init__(self, transforms):
  42. if not isinstance(transforms, list):
  43. raise TypeError('The transforms must be a list!')
  44. if len(transforms) < 1:
  45. raise ValueError('The length of transforms ' + \
  46. 'must be equal or larger than 1!')
  47. self.transforms = transforms
  48. self.batch_transforms = None
  49. self.use_mixup = False
  50. for t in self.transforms:
  51. if type(t).__name__ == 'MixupImage':
  52. self.use_mixup = True
  53. # 检查transforms里面的操作,目前支持PaddleX定义的或者是imgaug操作
  54. for op in self.transforms:
  55. if not isinstance(op, DetTransform):
  56. import imgaug.augmenters as iaa
  57. if not isinstance(op, iaa.Augmenter):
  58. raise Exception(
  59. "Elements in transforms should be defined in 'paddlex.det.transforms' or class of imgaug.augmenters.Augmenter, see docs here: https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/"
  60. )
  61. def __call__(self, im, im_info=None, label_info=None):
  62. """
  63. Args:
  64. im (str/np.ndarray): 图像路径/图像np.ndarray数据。
  65. im_info (dict): 存储与图像相关的信息,dict中的字段如下:
  66. - im_id (np.ndarray): 图像序列号,形状为(1,)。
  67. - image_shape (np.ndarray): 图像原始大小,形状为(2,),
  68. image_shape[0]为高,image_shape[1]为宽。
  69. - mixup (list): list为[im, im_info, label_info],分别对应
  70. 与当前图像进行mixup的图像np.ndarray数据、图像相关信息、标注框相关信息;
  71. 注意,当前epoch若无需进行mixup,则无该字段。
  72. label_info (dict): 存储与标注框相关的信息,dict中的字段如下:
  73. - gt_bbox (np.ndarray): 真实标注框坐标[x1, y1, x2, y2],形状为(n, 4),
  74. 其中n代表真实标注框的个数。
  75. - gt_class (np.ndarray): 每个真实标注框对应的类别序号,形状为(n, 1),
  76. 其中n代表真实标注框的个数。
  77. - gt_score (np.ndarray): 每个真实标注框对应的混合得分,形状为(n, 1),
  78. 其中n代表真实标注框的个数。
  79. - gt_poly (list): 每个真实标注框内的多边形分割区域,每个分割区域由点的x、y坐标组成,
  80. 长度为n,其中n代表真实标注框的个数。
  81. - is_crowd (np.ndarray): 每个真实标注框中是否是一组对象,形状为(n, 1),
  82. 其中n代表真实标注框的个数。
  83. - difficult (np.ndarray): 每个真实标注框中的对象是否为难识别对象,形状为(n, 1),
  84. 其中n代表真实标注框的个数。
  85. Returns:
  86. tuple: 根据网络所需字段所组成的tuple;
  87. 字段由transforms中的最后一个数据预处理操作决定。
  88. """
  89. def decode_image(im_file, im_info, label_info, input_channel=3):
  90. if im_info is None:
  91. im_info = dict()
  92. if isinstance(im_file, np.ndarray):
  93. if len(im_file.shape) != 3:
  94. raise Exception(
  95. "im should be 3-dimensions, but now is {}-dimensions".
  96. format(len(im_file.shape)))
  97. im = im_file
  98. else:
  99. try:
  100. if input_channel == 3:
  101. im = cv2.imread(im_file).astype('float32')
  102. else:
  103. im = cv2.imread(im_file,
  104. cv2.IMREAD_UNCHANGED).astype('float32')
  105. if im.ndim < 3:
  106. im = np.expand_dims(im, axis=-1)
  107. except:
  108. raise TypeError('Can\'t read The image file {}!'.format(
  109. im_file))
  110. im = im.astype('float32')
  111. if input_channel == 3:
  112. im = cv2.cvtColor(im, cv2.COLOR_BGR2RGB)
  113. # make default im_info with [h, w, 1]
  114. im_info['im_resize_info'] = np.array(
  115. [im.shape[0], im.shape[1], 1.], dtype=np.float32)
  116. im_info['image_shape'] = np.array([im.shape[0],
  117. im.shape[1]]).astype('int32')
  118. if not self.use_mixup:
  119. if 'mixup' in im_info:
  120. del im_info['mixup']
  121. # decode mixup image
  122. if 'mixup' in im_info:
  123. im_info['mixup'] = \
  124. decode_image(im_info['mixup'][0],
  125. im_info['mixup'][1],
  126. im_info['mixup'][2])
  127. if label_info is None:
  128. return (im, im_info)
  129. else:
  130. return (im, im_info, label_info)
  131. input_channel = getattr(self, 'input_channel', 3)
  132. outputs = decode_image(im, im_info, label_info, input_channel)
  133. im = outputs[0]
  134. im_info = outputs[1]
  135. if len(outputs) == 3:
  136. label_info = outputs[2]
  137. for op in self.transforms:
  138. if im is None:
  139. return None
  140. if isinstance(op, DetTransform):
  141. outputs = op(im, im_info, label_info)
  142. im = outputs[0]
  143. else:
  144. if im.shape[-1] != 3:
  145. raise Exception(
  146. "Only the 3-channel RGB image is supported in the imgaug operator, but recieved image channel is {}".
  147. format(im.shape[-1]))
  148. im = execute_imgaug(op, im)
  149. if label_info is not None:
  150. outputs = (im, im_info, label_info)
  151. else:
  152. outputs = (im, im_info)
  153. return outputs
  154. def add_augmenters(self, augmenters):
  155. if not isinstance(augmenters, list):
  156. raise Exception(
  157. "augmenters should be list type in func add_augmenters()")
  158. transform_names = [type(x).__name__ for x in self.transforms]
  159. for aug in augmenters:
  160. if type(aug).__name__ in transform_names:
  161. logging.error(
  162. "{} is already in ComposedTransforms, need to remove it from add_augmenters().".
  163. format(type(aug).__name__))
  164. self.transforms = augmenters + self.transforms
  165. class ResizeByShort(DetTransform):
  166. """根据图像的短边调整图像大小(resize)。
  167. 1. 获取图像的长边和短边长度。
  168. 2. 根据短边与short_size的比例,计算长边的目标长度,
  169. 此时高、宽的resize比例为short_size/原图短边长度。
  170. 若short_size为数组,则随机从该数组中挑选一个数值
  171. 作为short_size。
  172. 3. 如果max_size>0,调整resize比例:
  173. 如果长边的目标长度>max_size,则高、宽的resize比例为max_size/原图长边长度。
  174. 4. 根据调整大小的比例对图像进行resize。
  175. Args:
  176. short_size (int|list): 短边目标长度。默认为800。
  177. max_size (int): 长边目标长度的最大限制。默认为1333。
  178. Raises:
  179. TypeError: 形参数据类型不满足需求。
  180. """
  181. def __init__(self, short_size=800, max_size=1333):
  182. self.max_size = int(max_size)
  183. if not (isinstance(short_size, int) or isinstance(short_size, list)):
  184. raise TypeError(
  185. "Type of short_size is invalid. Must be Integer or List, now is {}".
  186. format(type(short_size)))
  187. self.short_size = short_size
  188. if not (isinstance(self.max_size, int)):
  189. raise TypeError("max_size: input type is invalid.")
  190. def __call__(self, im, im_info=None, label_info=None):
  191. """
  192. Args:
  193. im (numnp.ndarraypy): 图像np.ndarray数据。
  194. im_info (dict, 可选): 存储与图像相关的信息。
  195. label_info (dict, 可选): 存储与标注框相关的信息。
  196. Returns:
  197. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  198. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  199. 存储与标注框相关信息的字典。
  200. 其中,im_info更新字段为:
  201. - im_resize_info (np.ndarray): resize后的图像高、resize后的图像宽、resize后的图像相对原始图的缩放比例
  202. 三者组成的np.ndarray,形状为(3,)。
  203. Raises:
  204. TypeError: 形参数据类型不满足需求。
  205. ValueError: 数据长度不匹配。
  206. """
  207. if im_info is None:
  208. im_info = dict()
  209. if not isinstance(im, np.ndarray):
  210. raise TypeError("ResizeByShort: image type is not numpy.")
  211. if len(im.shape) != 3:
  212. raise ValueError('ResizeByShort: image is not 3-dimensional.')
  213. im_short_size = min(im.shape[0], im.shape[1])
  214. im_long_size = max(im.shape[0], im.shape[1])
  215. if isinstance(self.short_size, list):
  216. # Case for multi-scale training
  217. selected_size = random.choice(self.short_size)
  218. else:
  219. selected_size = self.short_size
  220. scale = float(selected_size) / im_short_size
  221. if self.max_size > 0 and np.round(scale *
  222. im_long_size) > self.max_size:
  223. scale = float(self.max_size) / float(im_long_size)
  224. resized_width = int(round(im.shape[1] * scale))
  225. resized_height = int(round(im.shape[0] * scale))
  226. im_resize_info = [resized_height, resized_width, scale]
  227. im = cv2.resize(
  228. im, (resized_width, resized_height),
  229. interpolation=cv2.INTER_LINEAR)
  230. if im.ndim < 3:
  231. im = np.expand_dims(im, axis=-1)
  232. im_info['im_resize_info'] = np.array(im_resize_info).astype(np.float32)
  233. if label_info is None:
  234. return (im, im_info)
  235. else:
  236. return (im, im_info, label_info)
  237. class Padding(DetTransform):
  238. """1.将图像的长和宽padding至coarsest_stride的倍数。如输入图像为[300, 640],
  239. `coarest_stride`为32,则由于300不为32的倍数,因此在图像最右和最下使用0值
  240. 进行padding,最终输出图像为[320, 640]。
  241. 2.或者,将图像的长和宽padding到target_size指定的shape,如输入的图像为[300,640],
  242. a. `target_size` = 960,在图像最右和最下使用0值进行padding,最终输出
  243. 图像为[960, 960]。
  244. b. `target_size` = [640, 960],在图像最右和最下使用0值进行padding,最终
  245. 输出图像为[640, 960]。
  246. 1. 如果coarsest_stride为1,target_size为None则直接返回。
  247. 2. 获取图像的高H、宽W。
  248. 3. 计算填充后图像的高H_new、宽W_new。
  249. 4. 构建大小为(H_new, W_new, 3)像素值为0的np.ndarray,
  250. 并将原图的np.ndarray粘贴于左上角。
  251. Args:
  252. coarsest_stride (int): 填充后的图像长、宽为该参数的倍数,默认为1。
  253. target_size (int|list|tuple): 填充后的图像长、宽,默认为None,coarset_stride优先级更高。
  254. Raises:
  255. TypeError: 形参`target_size`数据类型不满足需求。
  256. ValueError: 形参`target_size`为(list|tuple)时,长度不满足需求。
  257. """
  258. def __init__(self, coarsest_stride=1, target_size=None):
  259. self.coarsest_stride = coarsest_stride
  260. if target_size is not None:
  261. if not isinstance(target_size, int):
  262. if not isinstance(target_size, tuple) and not isinstance(
  263. target_size, list):
  264. raise TypeError(
  265. "Padding: Type of target_size must in (int|list|tuple)."
  266. )
  267. elif len(target_size) != 2:
  268. raise ValueError(
  269. "Padding: Length of target_size must equal 2.")
  270. self.target_size = target_size
  271. def __call__(self, im, im_info=None, label_info=None):
  272. """
  273. Args:
  274. im (numnp.ndarraypy): 图像np.ndarray数据。
  275. im_info (dict, 可选): 存储与图像相关的信息。
  276. label_info (dict, 可选): 存储与标注框相关的信息。
  277. Returns:
  278. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  279. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  280. 存储与标注框相关信息的字典。
  281. Raises:
  282. TypeError: 形参数据类型不满足需求。
  283. ValueError: 数据长度不匹配。
  284. ValueError: coarsest_stride,target_size需有且只有一个被指定。
  285. ValueError: target_size小于原图的大小。
  286. """
  287. if im_info is None:
  288. im_info = dict()
  289. if not isinstance(im, np.ndarray):
  290. raise TypeError("Padding: image type is not numpy.")
  291. if len(im.shape) != 3:
  292. raise ValueError('Padding: image is not 3-dimensional.')
  293. im_h, im_w, im_c = im.shape[:]
  294. if isinstance(self.target_size, int):
  295. padding_im_h = self.target_size
  296. padding_im_w = self.target_size
  297. elif isinstance(self.target_size, list) or isinstance(self.target_size,
  298. tuple):
  299. padding_im_w = self.target_size[0]
  300. padding_im_h = self.target_size[1]
  301. elif self.coarsest_stride > 0:
  302. padding_im_h = int(
  303. np.ceil(im_h / self.coarsest_stride) * self.coarsest_stride)
  304. padding_im_w = int(
  305. np.ceil(im_w / self.coarsest_stride) * self.coarsest_stride)
  306. else:
  307. raise ValueError(
  308. "coarsest_stridei(>1) or target_size(list|int) need setting in Padding transform"
  309. )
  310. pad_height = padding_im_h - im_h
  311. pad_width = padding_im_w - im_w
  312. if pad_height < 0 or pad_width < 0:
  313. raise ValueError(
  314. 'the size of image should be less than target_size, but the size of image ({}, {}), is larger than target_size ({}, {})'
  315. .format(im_w, im_h, padding_im_w, padding_im_h))
  316. padding_im = np.zeros(
  317. (padding_im_h, padding_im_w, im_c), dtype=np.float32)
  318. padding_im[:im_h, :im_w, :] = im
  319. if label_info is None:
  320. return (padding_im, im_info)
  321. else:
  322. return (padding_im, im_info, label_info)
  323. class Resize(DetTransform):
  324. """调整图像大小(resize)。
  325. - 当目标大小(target_size)类型为int时,根据插值方式,
  326. 将图像resize为[target_size, target_size]。
  327. - 当目标大小(target_size)类型为list或tuple时,根据插值方式,
  328. 将图像resize为target_size。
  329. 注意:当插值方式为“RANDOM”时,则随机选取一种插值方式进行resize。
  330. Args:
  331. target_size (int/list/tuple): 短边目标长度。默认为608。
  332. interp (str): resize的插值方式,与opencv的插值方式对应,取值范围为
  333. ['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4', 'RANDOM']。默认为"LINEAR"。
  334. Raises:
  335. TypeError: 形参数据类型不满足需求。
  336. ValueError: 插值方式不在['NEAREST', 'LINEAR', 'CUBIC',
  337. 'AREA', 'LANCZOS4', 'RANDOM']中。
  338. """
  339. # The interpolation mode
  340. interp_dict = {
  341. 'NEAREST': cv2.INTER_NEAREST,
  342. 'LINEAR': cv2.INTER_LINEAR,
  343. 'CUBIC': cv2.INTER_CUBIC,
  344. 'AREA': cv2.INTER_AREA,
  345. 'LANCZOS4': cv2.INTER_LANCZOS4
  346. }
  347. def __init__(self, target_size=608, interp='LINEAR'):
  348. self.interp = interp
  349. if not (interp == "RANDOM" or interp in self.interp_dict):
  350. raise ValueError("interp should be one of {}".format(
  351. self.interp_dict.keys()))
  352. if isinstance(target_size, list) or isinstance(target_size, tuple):
  353. if len(target_size) != 2:
  354. raise TypeError(
  355. 'when target is list or tuple, it should include 2 elements, but it is {}'
  356. .format(target_size))
  357. elif not isinstance(target_size, int):
  358. raise TypeError(
  359. "Type of target_size is invalid. Must be Integer or List or tuple, now is {}"
  360. .format(type(target_size)))
  361. self.target_size = target_size
  362. def __call__(self, im, im_info=None, label_info=None):
  363. """
  364. Args:
  365. im (np.ndarray): 图像np.ndarray数据。
  366. im_info (dict, 可选): 存储与图像相关的信息。
  367. label_info (dict, 可选): 存储与标注框相关的信息。
  368. Returns:
  369. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  370. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  371. 存储与标注框相关信息的字典。
  372. Raises:
  373. TypeError: 形参数据类型不满足需求。
  374. ValueError: 数据长度不匹配。
  375. """
  376. if im_info is None:
  377. im_info = dict()
  378. if not isinstance(im, np.ndarray):
  379. raise TypeError("Resize: image type is not numpy.")
  380. if len(im.shape) != 3:
  381. raise ValueError('Resize: image is not 3-dimensional.')
  382. if self.interp == "RANDOM":
  383. interp = random.choice(list(self.interp_dict.keys()))
  384. else:
  385. interp = self.interp
  386. im = resize(im, self.target_size, self.interp_dict[interp])
  387. if label_info is None:
  388. return (im, im_info)
  389. else:
  390. return (im, im_info, label_info)
  391. class RandomHorizontalFlip(DetTransform):
  392. """随机翻转图像、标注框、分割信息,模型训练时的数据增强操作。
  393. 1. 随机采样一个0-1之间的小数,当小数小于水平翻转概率时,
  394. 执行2-4步操作,否则直接返回。
  395. 2. 水平翻转图像。
  396. 3. 计算翻转后的真实标注框的坐标,更新label_info中的gt_bbox信息。
  397. 4. 计算翻转后的真实分割区域的坐标,更新label_info中的gt_poly信息。
  398. Args:
  399. prob (float): 随机水平翻转的概率。默认为0.5。
  400. Raises:
  401. TypeError: 形参数据类型不满足需求。
  402. """
  403. def __init__(self, prob=0.5):
  404. self.prob = prob
  405. if not isinstance(self.prob, float):
  406. raise TypeError("RandomHorizontalFlip: input type is invalid.")
  407. def __call__(self, im, im_info=None, label_info=None):
  408. """
  409. Args:
  410. im (np.ndarray): 图像np.ndarray数据。
  411. im_info (dict, 可选): 存储与图像相关的信息。
  412. label_info (dict, 可选): 存储与标注框相关的信息。
  413. Returns:
  414. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  415. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  416. 存储与标注框相关信息的字典。
  417. 其中,im_info更新字段为:
  418. - gt_bbox (np.ndarray): 水平翻转后的标注框坐标[x1, y1, x2, y2],形状为(n, 4),
  419. 其中n代表真实标注框的个数。
  420. - gt_poly (list): 水平翻转后的多边形分割区域的x、y坐标,长度为n,
  421. 其中n代表真实标注框的个数。
  422. Raises:
  423. TypeError: 形参数据类型不满足需求。
  424. ValueError: 数据长度不匹配。
  425. """
  426. if not isinstance(im, np.ndarray):
  427. raise TypeError(
  428. "RandomHorizontalFlip: image is not a numpy array.")
  429. if len(im.shape) != 3:
  430. raise ValueError(
  431. "RandomHorizontalFlip: image is not 3-dimensional.")
  432. if im_info is None or label_info is None:
  433. raise TypeError(
  434. 'Cannot do RandomHorizontalFlip! ' +
  435. 'Becasuse the im_info and label_info can not be None!')
  436. if 'gt_bbox' not in label_info:
  437. raise TypeError('Cannot do RandomHorizontalFlip! ' + \
  438. 'Becasuse gt_bbox is not in label_info!')
  439. image_shape = im_info['image_shape']
  440. gt_bbox = label_info['gt_bbox']
  441. height = image_shape[0]
  442. width = image_shape[1]
  443. if np.random.uniform(0, 1) < self.prob:
  444. im = horizontal_flip(im)
  445. if gt_bbox.shape[0] == 0:
  446. if label_info is None:
  447. return (im, im_info)
  448. else:
  449. return (im, im_info, label_info)
  450. label_info['gt_bbox'] = box_horizontal_flip(gt_bbox, width)
  451. if 'gt_poly' in label_info and \
  452. len(label_info['gt_poly']) != 0:
  453. label_info['gt_poly'] = segms_horizontal_flip(
  454. label_info['gt_poly'], height, width)
  455. if label_info is None:
  456. return (im, im_info)
  457. else:
  458. return (im, im_info, label_info)
  459. class Normalize(DetTransform):
  460. """对图像进行标准化。
  461. 1. 归一化图像到到区间[0.0, 1.0]。
  462. 2. 对图像进行减均值除以标准差操作。
  463. Args:
  464. mean (list): 图像数据集的均值。默认为[0.485, 0.456, 0.406]。
  465. std (list): 图像数据集的标准差。默认为[0.229, 0.224, 0.225]。
  466. Raises:
  467. TypeError: 形参数据类型不满足需求。
  468. """
  469. def __init__(self, mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]):
  470. self.mean = mean
  471. self.std = std
  472. if not (isinstance(self.mean, list) and isinstance(self.std, list)):
  473. raise TypeError("NormalizeImage: input type is invalid.")
  474. from functools import reduce
  475. if reduce(lambda x, y: x * y, self.std) == 0:
  476. raise TypeError('NormalizeImage: std is invalid!')
  477. def __call__(self, im, im_info=None, label_info=None):
  478. """
  479. Args:
  480. im (numnp.ndarraypy): 图像np.ndarray数据。
  481. im_info (dict, 可选): 存储与图像相关的信息。
  482. label_info (dict, 可选): 存储与标注框相关的信息。
  483. Returns:
  484. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  485. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  486. 存储与标注框相关信息的字典。
  487. """
  488. mean = np.array(self.mean)[np.newaxis, np.newaxis, :]
  489. std = np.array(self.std)[np.newaxis, np.newaxis, :]
  490. min_val = [0] * im.shape[-1]
  491. max_val = [255] * im.shape[-1]
  492. im = normalize(im, mean, std, min_val, max_val)
  493. if label_info is None:
  494. return (im, im_info)
  495. else:
  496. return (im, im_info, label_info)
  497. class RandomDistort(DetTransform):
  498. """以一定的概率对图像进行随机像素内容变换,模型训练时的数据增强操作
  499. 1. 对变换的操作顺序进行随机化操作。
  500. 2. 按照1中的顺序以一定的概率在范围[-range, range]对图像进行随机像素内容变换。
  501. Args:
  502. brightness_range (float): 明亮度因子的范围。默认为0.5。
  503. brightness_prob (float): 随机调整明亮度的概率。默认为0.5。
  504. contrast_range (float): 对比度因子的范围。默认为0.5。
  505. contrast_prob (float): 随机调整对比度的概率。默认为0.5。
  506. saturation_range (float): 饱和度因子的范围。默认为0.5。
  507. saturation_prob (float): 随机调整饱和度的概率。默认为0.5。
  508. hue_range (int): 色调因子的范围。默认为18。
  509. hue_prob (float): 随机调整色调的概率。默认为0.5。
  510. """
  511. def __init__(self,
  512. brightness_range=0.5,
  513. brightness_prob=0.5,
  514. contrast_range=0.5,
  515. contrast_prob=0.5,
  516. saturation_range=0.5,
  517. saturation_prob=0.5,
  518. hue_range=18,
  519. hue_prob=0.5):
  520. self.brightness_range = brightness_range
  521. self.brightness_prob = brightness_prob
  522. self.contrast_range = contrast_range
  523. self.contrast_prob = contrast_prob
  524. self.saturation_range = saturation_range
  525. self.saturation_prob = saturation_prob
  526. self.hue_range = hue_range
  527. self.hue_prob = hue_prob
  528. def __call__(self, im, im_info=None, label_info=None):
  529. """
  530. Args:
  531. im (np.ndarray): 图像np.ndarray数据。
  532. im_info (dict, 可选): 存储与图像相关的信息。
  533. label_info (dict, 可选): 存储与标注框相关的信息。
  534. Returns:
  535. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  536. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  537. 存储与标注框相关信息的字典。
  538. """
  539. if im.shape[-1] != 3:
  540. raise Exception(
  541. "Only the 3-channel RGB image is supported in the RandomDistort operator, but recieved image channel is {}".
  542. format(im.shape[-1]))
  543. brightness_lower = 1 - self.brightness_range
  544. brightness_upper = 1 + self.brightness_range
  545. contrast_lower = 1 - self.contrast_range
  546. contrast_upper = 1 + self.contrast_range
  547. saturation_lower = 1 - self.saturation_range
  548. saturation_upper = 1 + self.saturation_range
  549. hue_lower = -self.hue_range
  550. hue_upper = self.hue_range
  551. ops = [brightness, contrast, saturation, hue]
  552. random.shuffle(ops)
  553. params_dict = {
  554. 'brightness': {
  555. 'brightness_lower': brightness_lower,
  556. 'brightness_upper': brightness_upper
  557. },
  558. 'contrast': {
  559. 'contrast_lower': contrast_lower,
  560. 'contrast_upper': contrast_upper
  561. },
  562. 'saturation': {
  563. 'saturation_lower': saturation_lower,
  564. 'saturation_upper': saturation_upper
  565. },
  566. 'hue': {
  567. 'hue_lower': hue_lower,
  568. 'hue_upper': hue_upper
  569. }
  570. }
  571. prob_dict = {
  572. 'brightness': self.brightness_prob,
  573. 'contrast': self.contrast_prob,
  574. 'saturation': self.saturation_prob,
  575. 'hue': self.hue_prob
  576. }
  577. for id in range(4):
  578. params = params_dict[ops[id].__name__]
  579. prob = prob_dict[ops[id].__name__]
  580. params['im'] = im
  581. if np.random.uniform(0, 1) < prob:
  582. im = ops[id](**params)
  583. im = im.astype('float32')
  584. if label_info is None:
  585. return (im, im_info)
  586. else:
  587. return (im, im_info, label_info)
  588. class MixupImage(DetTransform):
  589. """对图像进行mixup操作,模型训练时的数据增强操作,目前仅YOLOv3模型支持该transform。
  590. 当label_info中不存在mixup字段时,直接返回,否则进行下述操作:
  591. 1. 从随机beta分布中抽取出随机因子factor。
  592. 2.
  593. - 当factor>=1.0时,去除label_info中的mixup字段,直接返回。
  594. - 当factor<=0.0时,直接返回label_info中的mixup字段,并在label_info中去除该字段。
  595. - 其余情况,执行下述操作:
  596. (1)原图像乘以factor,mixup图像乘以(1-factor),叠加2个结果。
  597. (2)拼接原图像标注框和mixup图像标注框。
  598. (3)拼接原图像标注框类别和mixup图像标注框类别。
  599. (4)原图像标注框混合得分乘以factor,mixup图像标注框混合得分乘以(1-factor),叠加2个结果。
  600. 3. 更新im_info中的image_shape信息。
  601. Args:
  602. alpha (float): 随机beta分布的下限。默认为1.5。
  603. beta (float): 随机beta分布的上限。默认为1.5。
  604. mixup_epoch (int): 在前mixup_epoch轮使用mixup增强操作;当该参数为-1时,该策略不会生效。
  605. 默认为-1。
  606. Raises:
  607. ValueError: 数据长度不匹配。
  608. """
  609. def __init__(self, alpha=1.5, beta=1.5, mixup_epoch=-1):
  610. self.alpha = alpha
  611. self.beta = beta
  612. if self.alpha <= 0.0:
  613. raise ValueError("alpha shold be positive in MixupImage")
  614. if self.beta <= 0.0:
  615. raise ValueError("beta shold be positive in MixupImage")
  616. self.mixup_epoch = mixup_epoch
  617. def _mixup_img(self, img1, img2, factor):
  618. h = max(img1.shape[0], img2.shape[0])
  619. w = max(img1.shape[1], img2.shape[1])
  620. img = np.zeros((h, w, img1.shape[2]), 'float32')
  621. img[:img1.shape[0], :img1.shape[1], :] = \
  622. img1.astype('float32') * factor
  623. img[:img2.shape[0], :img2.shape[1], :] += \
  624. img2.astype('float32') * (1.0 - factor)
  625. return img.astype('float32')
  626. def __call__(self, im, im_info=None, label_info=None):
  627. """
  628. Args:
  629. im (np.ndarray): 图像np.ndarray数据。
  630. im_info (dict, 可选): 存储与图像相关的信息。
  631. label_info (dict, 可选): 存储与标注框相关的信息。
  632. Returns:
  633. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  634. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  635. 存储与标注框相关信息的字典。
  636. 其中,im_info更新字段为:
  637. - image_shape (np.ndarray): mixup后的图像高、宽二者组成的np.ndarray,形状为(2,)。
  638. im_info删除的字段:
  639. - mixup (list): 与当前字段进行mixup的图像相关信息。
  640. label_info更新字段为:
  641. - gt_bbox (np.ndarray): mixup后真实标注框坐标,形状为(n, 4),
  642. 其中n代表真实标注框的个数。
  643. - gt_class (np.ndarray): mixup后每个真实标注框对应的类别序号,形状为(n, 1),
  644. 其中n代表真实标注框的个数。
  645. - gt_score (np.ndarray): mixup后每个真实标注框对应的混合得分,形状为(n, 1),
  646. 其中n代表真实标注框的个数。
  647. Raises:
  648. TypeError: 形参数据类型不满足需求。
  649. """
  650. if im_info is None:
  651. raise TypeError('Cannot do MixupImage! ' +
  652. 'Becasuse the im_info can not be None!')
  653. if 'mixup' not in im_info:
  654. if label_info is None:
  655. return (im, im_info)
  656. else:
  657. return (im, im_info, label_info)
  658. factor = np.random.beta(self.alpha, self.beta)
  659. factor = max(0.0, min(1.0, factor))
  660. if im_info['epoch'] > self.mixup_epoch \
  661. or factor >= 1.0:
  662. im_info.pop('mixup')
  663. if label_info is None:
  664. return (im, im_info)
  665. else:
  666. return (im, im_info, label_info)
  667. if factor <= 0.0:
  668. return im_info.pop('mixup')
  669. im = self._mixup_img(im, im_info['mixup'][0], factor)
  670. if label_info is None:
  671. raise TypeError('Cannot do MixupImage! ' +
  672. 'Becasuse the label_info can not be None!')
  673. if 'gt_bbox' not in label_info or \
  674. 'gt_class' not in label_info or \
  675. 'gt_score' not in label_info:
  676. raise TypeError('Cannot do MixupImage! ' + \
  677. 'Becasuse gt_bbox/gt_class/gt_score is not in label_info!')
  678. gt_bbox1 = label_info['gt_bbox']
  679. gt_bbox2 = im_info['mixup'][2]['gt_bbox']
  680. gt_class1 = label_info['gt_class']
  681. gt_class2 = im_info['mixup'][2]['gt_class']
  682. gt_score1 = label_info['gt_score']
  683. gt_score2 = im_info['mixup'][2]['gt_score']
  684. if 'gt_poly' in label_info:
  685. gt_poly1 = label_info['gt_poly']
  686. gt_poly2 = im_info['mixup'][2]['gt_poly']
  687. is_crowd1 = label_info['is_crowd']
  688. is_crowd2 = im_info['mixup'][2]['is_crowd']
  689. if 0 not in gt_class1 and 0 not in gt_class2:
  690. gt_bbox = np.concatenate((gt_bbox1, gt_bbox2), axis=0)
  691. gt_class = np.concatenate((gt_class1, gt_class2), axis=0)
  692. gt_score = np.concatenate(
  693. (gt_score1 * factor, gt_score2 * (1. - factor)), axis=0)
  694. if 'gt_poly' in label_info:
  695. label_info['gt_poly'] = gt_poly1 + gt_poly2
  696. is_crowd = np.concatenate((is_crowd1, is_crowd2), axis=0)
  697. elif 0 in gt_class1:
  698. gt_bbox = gt_bbox2
  699. gt_class = gt_class2
  700. gt_score = gt_score2 * (1. - factor)
  701. if 'gt_poly' in label_info:
  702. label_info['gt_poly'] = gt_poly2
  703. is_crowd = is_crowd2
  704. else:
  705. gt_bbox = gt_bbox1
  706. gt_class = gt_class1
  707. gt_score = gt_score1 * factor
  708. if 'gt_poly' in label_info:
  709. label_info['gt_poly'] = gt_poly1
  710. is_crowd = is_crowd1
  711. label_info['gt_bbox'] = gt_bbox
  712. label_info['gt_score'] = gt_score
  713. label_info['gt_class'] = gt_class
  714. label_info['is_crowd'] = is_crowd
  715. im_info['image_shape'] = np.array([im.shape[0],
  716. im.shape[1]]).astype('int32')
  717. im_info.pop('mixup')
  718. if label_info is None:
  719. return (im, im_info)
  720. else:
  721. return (im, im_info, label_info)
  722. class RandomExpand(DetTransform):
  723. """随机扩张图像,模型训练时的数据增强操作。
  724. 1. 随机选取扩张比例(扩张比例大于1时才进行扩张)。
  725. 2. 计算扩张后图像大小。
  726. 3. 初始化像素值为输入填充值的图像,并将原图像随机粘贴于该图像上。
  727. 4. 根据原图像粘贴位置换算出扩张后真实标注框的位置坐标。
  728. 5. 根据原图像粘贴位置换算出扩张后真实分割区域的位置坐标。
  729. Args:
  730. ratio (float): 图像扩张的最大比例。默认为4.0。
  731. prob (float): 随机扩张的概率。默认为0.5。
  732. fill_value (list): 扩张图像的初始填充值(0-255)。默认为[123.675, 116.28, 103.53]。
  733. """
  734. def __init__(self,
  735. ratio=4.,
  736. prob=0.5,
  737. fill_value=[123.675, 116.28, 103.53]):
  738. super(RandomExpand, self).__init__()
  739. assert ratio > 1.01, "expand ratio must be larger than 1.01"
  740. self.ratio = ratio
  741. self.prob = prob
  742. assert isinstance(fill_value, Sequence), \
  743. "fill value must be sequence"
  744. if not isinstance(fill_value, tuple):
  745. fill_value = tuple(fill_value)
  746. self.fill_value = fill_value
  747. def __call__(self, im, im_info=None, label_info=None):
  748. """
  749. Args:
  750. im (np.ndarray): 图像np.ndarray数据。
  751. im_info (dict, 可选): 存储与图像相关的信息。
  752. label_info (dict, 可选): 存储与标注框相关的信息。
  753. Returns:
  754. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  755. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  756. 存储与标注框相关信息的字典。
  757. 其中,im_info更新字段为:
  758. - image_shape (np.ndarray): 扩张后的图像高、宽二者组成的np.ndarray,形状为(2,)。
  759. label_info更新字段为:
  760. - gt_bbox (np.ndarray): 随机扩张后真实标注框坐标,形状为(n, 4),
  761. 其中n代表真实标注框的个数。
  762. - gt_class (np.ndarray): 随机扩张后每个真实标注框对应的类别序号,形状为(n, 1),
  763. 其中n代表真实标注框的个数。
  764. Raises:
  765. TypeError: 形参数据类型不满足需求。
  766. """
  767. if im_info is None or label_info is None:
  768. raise TypeError(
  769. 'Cannot do RandomExpand! ' +
  770. 'Becasuse the im_info and label_info can not be None!')
  771. if 'gt_bbox' not in label_info or \
  772. 'gt_class' not in label_info:
  773. raise TypeError('Cannot do RandomExpand! ' + \
  774. 'Becasuse gt_bbox/gt_class is not in label_info!')
  775. if np.random.uniform(0., 1.) > self.prob:
  776. return (im, im_info, label_info)
  777. if 'gt_class' in label_info and 0 in label_info['gt_class']:
  778. return (im, im_info, label_info)
  779. image_shape = im_info['image_shape']
  780. height = int(image_shape[0])
  781. width = int(image_shape[1])
  782. expand_ratio = np.random.uniform(1., self.ratio)
  783. h = int(height * expand_ratio)
  784. w = int(width * expand_ratio)
  785. if not h > height or not w > width:
  786. return (im, im_info, label_info)
  787. y = np.random.randint(0, h - height)
  788. x = np.random.randint(0, w - width)
  789. canvas = np.ones((h, w, 3), dtype=np.float32)
  790. canvas *= np.array(self.fill_value, dtype=np.float32)
  791. canvas[y:y + height, x:x + width, :] = im
  792. im_info['image_shape'] = np.array([h, w]).astype('int32')
  793. if 'gt_bbox' in label_info and len(label_info['gt_bbox']) > 0:
  794. label_info['gt_bbox'] += np.array([x, y] * 2, dtype=np.float32)
  795. if 'gt_poly' in label_info and len(label_info['gt_poly']) > 0:
  796. label_info['gt_poly'] = expand_segms(label_info['gt_poly'], x, y,
  797. height, width, expand_ratio)
  798. return (canvas, im_info, label_info)
  799. class RandomCrop(DetTransform):
  800. """随机裁剪图像。
  801. 1. 若allow_no_crop为True,则在thresholds加入’no_crop’。
  802. 2. 随机打乱thresholds。
  803. 3. 遍历thresholds中各元素:
  804. (1) 如果当前thresh为’no_crop’,则返回原始图像和标注信息。
  805. (2) 随机取出aspect_ratio和scaling中的值并由此计算出候选裁剪区域的高、宽、起始点。
  806. (3) 计算真实标注框与候选裁剪区域IoU,若全部真实标注框的IoU都小于thresh,则继续第3步。
  807. (4) 如果cover_all_box为True且存在真实标注框的IoU小于thresh,则继续第3步。
  808. (5) 筛选出位于候选裁剪区域内的真实标注框,若有效框的个数为0,则继续第3步,否则进行第4步。
  809. 4. 换算有效真值标注框相对候选裁剪区域的位置坐标。
  810. 5. 换算有效分割区域相对候选裁剪区域的位置坐标。
  811. Args:
  812. aspect_ratio (list): 裁剪后短边缩放比例的取值范围,以[min, max]形式表示。默认值为[.5, 2.]。
  813. thresholds (list): 判断裁剪候选区域是否有效所需的IoU阈值取值列表。默认值为[.0, .1, .3, .5, .7, .9]。
  814. scaling (list): 裁剪面积相对原面积的取值范围,以[min, max]形式表示。默认值为[.3, 1.]。
  815. num_attempts (int): 在放弃寻找有效裁剪区域前尝试的次数。默认值为50。
  816. allow_no_crop (bool): 是否允许未进行裁剪。默认值为True。
  817. cover_all_box (bool): 是否要求所有的真实标注框都必须在裁剪区域内。默认值为False。
  818. """
  819. def __init__(self,
  820. aspect_ratio=[.5, 2.],
  821. thresholds=[.0, .1, .3, .5, .7, .9],
  822. scaling=[.3, 1.],
  823. num_attempts=50,
  824. allow_no_crop=True,
  825. cover_all_box=False):
  826. self.aspect_ratio = aspect_ratio
  827. self.thresholds = thresholds
  828. self.scaling = scaling
  829. self.num_attempts = num_attempts
  830. self.allow_no_crop = allow_no_crop
  831. self.cover_all_box = cover_all_box
  832. def __call__(self, im, im_info=None, label_info=None):
  833. """
  834. Args:
  835. im (np.ndarray): 图像np.ndarray数据。
  836. im_info (dict, 可选): 存储与图像相关的信息。
  837. label_info (dict, 可选): 存储与标注框相关的信息。
  838. Returns:
  839. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  840. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  841. 存储与标注框相关信息的字典。
  842. 其中,im_info更新字段为:
  843. - image_shape (np.ndarray): 扩裁剪的图像高、宽二者组成的np.ndarray,形状为(2,)。
  844. label_info更新字段为:
  845. - gt_bbox (np.ndarray): 随机裁剪后真实标注框坐标,形状为(n, 4),
  846. 其中n代表真实标注框的个数。
  847. - gt_class (np.ndarray): 随机裁剪后每个真实标注框对应的类别序号,形状为(n, 1),
  848. 其中n代表真实标注框的个数。
  849. - gt_score (np.ndarray): 随机裁剪后每个真实标注框对应的混合得分,形状为(n, 1),
  850. 其中n代表真实标注框的个数。
  851. Raises:
  852. TypeError: 形参数据类型不满足需求。
  853. """
  854. if im_info is None or label_info is None:
  855. raise TypeError(
  856. 'Cannot do RandomCrop! ' +
  857. 'Becasuse the im_info and label_info can not be None!')
  858. if 'gt_bbox' not in label_info or \
  859. 'gt_class' not in label_info:
  860. raise TypeError('Cannot do RandomCrop! ' + \
  861. 'Becasuse gt_bbox/gt_class is not in label_info!')
  862. if len(label_info['gt_bbox']) == 0:
  863. return (im, im_info, label_info)
  864. if 'gt_class' in label_info and 0 in label_info['gt_class']:
  865. return (im, im_info, label_info)
  866. image_shape = im_info['image_shape']
  867. w = image_shape[1]
  868. h = image_shape[0]
  869. gt_bbox = label_info['gt_bbox']
  870. thresholds = list(self.thresholds)
  871. if self.allow_no_crop:
  872. thresholds.append('no_crop')
  873. np.random.shuffle(thresholds)
  874. for thresh in thresholds:
  875. if thresh == 'no_crop':
  876. return (im, im_info, label_info)
  877. found = False
  878. for i in range(self.num_attempts):
  879. scale = np.random.uniform(*self.scaling)
  880. min_ar, max_ar = self.aspect_ratio
  881. aspect_ratio = np.random.uniform(
  882. max(min_ar, scale**2), min(max_ar, scale**-2))
  883. crop_h = int(h * scale / np.sqrt(aspect_ratio))
  884. crop_w = int(w * scale * np.sqrt(aspect_ratio))
  885. crop_y = np.random.randint(0, h - crop_h)
  886. crop_x = np.random.randint(0, w - crop_w)
  887. crop_box = [crop_x, crop_y, crop_x + crop_w, crop_y + crop_h]
  888. iou = iou_matrix(
  889. gt_bbox, np.array(
  890. [crop_box], dtype=np.float32))
  891. if iou.max() < thresh:
  892. continue
  893. if self.cover_all_box and iou.min() < thresh:
  894. continue
  895. cropped_box, valid_ids = crop_box_with_center_constraint(
  896. gt_bbox, np.array(
  897. crop_box, dtype=np.float32))
  898. if valid_ids.size > 0:
  899. found = True
  900. break
  901. if found:
  902. if 'gt_poly' in label_info and len(label_info['gt_poly']) > 0:
  903. crop_polys = crop_segms(
  904. label_info['gt_poly'],
  905. valid_ids,
  906. np.array(
  907. crop_box, dtype=np.int64),
  908. h,
  909. w)
  910. if [] in crop_polys:
  911. delete_id = list()
  912. valid_polys = list()
  913. for id, crop_poly in enumerate(crop_polys):
  914. if crop_poly == []:
  915. delete_id.append(id)
  916. else:
  917. valid_polys.append(crop_poly)
  918. valid_ids = np.delete(valid_ids, delete_id)
  919. if len(valid_polys) == 0:
  920. return (im, im_info, label_info)
  921. label_info['gt_poly'] = valid_polys
  922. else:
  923. label_info['gt_poly'] = crop_polys
  924. im = crop_image(im, crop_box)
  925. label_info['gt_bbox'] = np.take(cropped_box, valid_ids, axis=0)
  926. label_info['gt_class'] = np.take(
  927. label_info['gt_class'], valid_ids, axis=0)
  928. im_info['image_shape'] = np.array(
  929. [crop_box[3] - crop_box[1],
  930. crop_box[2] - crop_box[0]]).astype('int32')
  931. if 'gt_score' in label_info:
  932. label_info['gt_score'] = np.take(
  933. label_info['gt_score'], valid_ids, axis=0)
  934. if 'is_crowd' in label_info:
  935. label_info['is_crowd'] = np.take(
  936. label_info['is_crowd'], valid_ids, axis=0)
  937. return (im, im_info, label_info)
  938. return (im, im_info, label_info)
  939. class CLAHE(DetTransform):
  940. """对图像进行对比度增强。
  941. Args:
  942. clip_limit (int|float): 颜色对比度的阈值,默认值为2.。
  943. tile_grid_size (list|tuple): 进行像素均衡化的网格大小。默认值为(8, 8)。
  944. Raises:
  945. TypeError: 形参数据类型不满足需求。
  946. """
  947. def __init__(self, clip_limit=2., tile_grid_size=(8, 8)):
  948. self.clip_limit = clip_limit
  949. self.tile_grid_size = tile_grid_size
  950. def __call__(self, im, im_info=None, label_info=None):
  951. """
  952. Args:
  953. im (numnp.ndarraypy): 图像np.ndarray数据。
  954. im_info (dict, 可选): 存储与图像相关的信息。
  955. label_info (dict, 可选): 存储与标注框相关的信息。
  956. Returns:
  957. tuple: 当label_info为空时,返回的tuple为(im, im_info),分别对应图像np.ndarray数据、存储与图像相关信息的字典;
  958. 当label_info不为空时,返回的tuple为(im, im_info, label_info),分别对应图像np.ndarray数据、
  959. 存储与标注框相关信息的字典。
  960. """
  961. if im.shape[-1] != 1:
  962. raise Exception(
  963. "Only the one-channel image is supported in the CLAHE operator, but recieved image channel is {}".
  964. format(im.shape[-1]))
  965. clahe = cv2.createCLAHE(
  966. clipLimit=self.clip_limit, tileGridSize=self.tile_grid_size)
  967. im = clahe.apply(im).astype(im.dtype)
  968. if label_info is None:
  969. return (im, im_info)
  970. else:
  971. return (im, im_info, label_info)
  972. class ArrangeFasterRCNN(DetTransform):
  973. """获取FasterRCNN模型训练/验证/预测所需信息。
  974. Args:
  975. mode (str): 指定数据用于何种用途,取值范围为['train', 'eval', 'test', 'quant']。
  976. Raises:
  977. ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
  978. """
  979. def __init__(self, mode=None):
  980. if mode not in ['train', 'eval', 'test', 'quant']:
  981. raise ValueError(
  982. "mode must be in ['train', 'eval', 'test', 'quant']!")
  983. self.mode = mode
  984. def __call__(self, im, im_info=None, label_info=None):
  985. """
  986. Args:
  987. im (np.ndarray): 图像np.ndarray数据。
  988. im_info (dict, 可选): 存储与图像相关的信息。
  989. label_info (dict, 可选): 存储与标注框相关的信息。
  990. Returns:
  991. tuple: 当mode为'train'时,返回(im, im_resize_info, gt_bbox, gt_class, is_crowd),分别对应
  992. 图像np.ndarray数据、图像相当对于原图的resize信息、真实标注框、真实标注框对应的类别、真实标注框内是否是一组对象;
  993. 当mode为'eval'时,返回(im, im_resize_info, im_id, im_shape, gt_bbox, gt_class, is_difficult),
  994. 分别对应图像np.ndarray数据、图像相当对于原图的resize信息、图像id、图像大小信息、真实标注框、真实标注框对应的类别、
  995. 真实标注框是否为难识别对象;当mode为'test'或'quant'时,返回(im, im_resize_info, im_shape),分别对应图像np.ndarray数据、
  996. 图像相当对于原图的resize信息、图像大小信息。
  997. Raises:
  998. TypeError: 形参数据类型不满足需求。
  999. ValueError: 数据长度不匹配。
  1000. """
  1001. im = permute(im, False)
  1002. if self.mode == 'train':
  1003. if im_info is None or label_info is None:
  1004. raise TypeError(
  1005. 'Cannot do ArrangeFasterRCNN! ' +
  1006. 'Becasuse the im_info and label_info can not be None!')
  1007. if len(label_info['gt_bbox']) != len(label_info['gt_class']):
  1008. raise ValueError("gt num mismatch: bbox and class.")
  1009. im_resize_info = im_info['im_resize_info']
  1010. gt_bbox = label_info['gt_bbox']
  1011. gt_class = label_info['gt_class']
  1012. is_crowd = label_info['is_crowd']
  1013. outputs = (im, im_resize_info, gt_bbox, gt_class, is_crowd)
  1014. elif self.mode == 'eval':
  1015. if im_info is None or label_info is None:
  1016. raise TypeError(
  1017. 'Cannot do ArrangeFasterRCNN! ' +
  1018. 'Becasuse the im_info and label_info can not be None!')
  1019. im_resize_info = im_info['im_resize_info']
  1020. im_id = im_info['im_id']
  1021. im_shape = np.array(
  1022. (im_info['image_shape'][0], im_info['image_shape'][1], 1),
  1023. dtype=np.float32)
  1024. gt_bbox = label_info['gt_bbox']
  1025. gt_class = label_info['gt_class']
  1026. is_difficult = label_info['difficult']
  1027. outputs = (im, im_resize_info, im_id, im_shape, gt_bbox, gt_class,
  1028. is_difficult)
  1029. else:
  1030. if im_info is None:
  1031. raise TypeError('Cannot do ArrangeFasterRCNN! ' +
  1032. 'Becasuse the im_info can not be None!')
  1033. im_resize_info = im_info['im_resize_info']
  1034. im_shape = np.array(
  1035. (im_info['image_shape'][0], im_info['image_shape'][1], 1),
  1036. dtype=np.float32)
  1037. outputs = (im, im_resize_info, im_shape)
  1038. return outputs
  1039. class ArrangeMaskRCNN(DetTransform):
  1040. """获取MaskRCNN模型训练/验证/预测所需信息。
  1041. Args:
  1042. mode (str): 指定数据用于何种用途,取值范围为['train', 'eval', 'test', 'quant']。
  1043. Raises:
  1044. ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
  1045. """
  1046. def __init__(self, mode=None):
  1047. if mode not in ['train', 'eval', 'test', 'quant']:
  1048. raise ValueError(
  1049. "mode must be in ['train', 'eval', 'test', 'quant']!")
  1050. self.mode = mode
  1051. def __call__(self, im, im_info=None, label_info=None):
  1052. """
  1053. Args:
  1054. im (np.ndarray): 图像np.ndarray数据。
  1055. im_info (dict, 可选): 存储与图像相关的信息。
  1056. label_info (dict, 可选): 存储与标注框相关的信息。
  1057. Returns:
  1058. tuple: 当mode为'train'时,返回(im, im_resize_info, gt_bbox, gt_class, is_crowd, gt_masks),分别对应
  1059. 图像np.ndarray数据、图像相当对于原图的resize信息、真实标注框、真实标注框对应的类别、真实标注框内是否是一组对象、
  1060. 真实分割区域;当mode为'eval'时,返回(im, im_resize_info, im_id, im_shape),分别对应图像np.ndarray数据、
  1061. 图像相当对于原图的resize信息、图像id、图像大小信息;当mode为'test'或'quant'时,返回(im, im_resize_info, im_shape),
  1062. 分别对应图像np.ndarray数据、图像相当对于原图的resize信息、图像大小信息。
  1063. Raises:
  1064. TypeError: 形参数据类型不满足需求。
  1065. ValueError: 数据长度不匹配。
  1066. """
  1067. im = permute(im, False)
  1068. if self.mode == 'train':
  1069. if im_info is None or label_info is None:
  1070. raise TypeError(
  1071. 'Cannot do ArrangeTrainMaskRCNN! ' +
  1072. 'Becasuse the im_info and label_info can not be None!')
  1073. if len(label_info['gt_bbox']) != len(label_info['gt_class']):
  1074. raise ValueError("gt num mismatch: bbox and class.")
  1075. im_resize_info = im_info['im_resize_info']
  1076. gt_bbox = label_info['gt_bbox']
  1077. gt_class = label_info['gt_class']
  1078. is_crowd = label_info['is_crowd']
  1079. assert 'gt_poly' in label_info
  1080. segms = label_info['gt_poly']
  1081. if len(segms) != 0:
  1082. assert len(segms) == is_crowd.shape[0]
  1083. gt_masks = []
  1084. valid = True
  1085. for i in range(len(segms)):
  1086. segm = segms[i]
  1087. gt_segm = []
  1088. if is_crowd[i]:
  1089. gt_segm.append([[0, 0]])
  1090. else:
  1091. for poly in segm:
  1092. if len(poly) == 0:
  1093. valid = False
  1094. break
  1095. gt_segm.append(np.array(poly).reshape(-1, 2))
  1096. if (not valid) or len(gt_segm) == 0:
  1097. break
  1098. gt_masks.append(gt_segm)
  1099. outputs = (im, im_resize_info, gt_bbox, gt_class, is_crowd,
  1100. gt_masks)
  1101. else:
  1102. if im_info is None:
  1103. raise TypeError('Cannot do ArrangeMaskRCNN! ' +
  1104. 'Becasuse the im_info can not be None!')
  1105. im_resize_info = im_info['im_resize_info']
  1106. im_shape = np.array(
  1107. (im_info['image_shape'][0], im_info['image_shape'][1], 1),
  1108. dtype=np.float32)
  1109. if self.mode == 'eval':
  1110. im_id = im_info['im_id']
  1111. outputs = (im, im_resize_info, im_id, im_shape)
  1112. else:
  1113. outputs = (im, im_resize_info, im_shape)
  1114. return outputs
  1115. class ArrangeYOLOv3(DetTransform):
  1116. """获取YOLOv3模型训练/验证/预测所需信息。
  1117. Args:
  1118. mode (str): 指定数据用于何种用途,取值范围为['train', 'eval', 'test', 'quant']。
  1119. Raises:
  1120. ValueError: mode的取值不在['train', 'eval', 'test', 'quant']之内。
  1121. """
  1122. def __init__(self, mode=None):
  1123. if mode not in ['train', 'eval', 'test', 'quant']:
  1124. raise ValueError(
  1125. "mode must be in ['train', 'eval', 'test', 'quant']!")
  1126. self.mode = mode
  1127. def __call__(self, im, im_info=None, label_info=None):
  1128. """
  1129. Args:
  1130. im (np.ndarray): 图像np.ndarray数据。
  1131. im_info (dict, 可选): 存储与图像相关的信息。
  1132. label_info (dict, 可选): 存储与标注框相关的信息。
  1133. Returns:
  1134. tuple: 当mode为'train'时,返回(im, gt_bbox, gt_class, gt_score, im_shape),分别对应
  1135. 图像np.ndarray数据、真实标注框、真实标注框对应的类别、真实标注框混合得分、图像大小信息;
  1136. 当mode为'eval'时,返回(im, im_shape, im_id, gt_bbox, gt_class, difficult),
  1137. 分别对应图像np.ndarray数据、图像大小信息、图像id、真实标注框、真实标注框对应的类别、
  1138. 真实标注框是否为难识别对象;当mode为'test'或'quant'时,返回(im, im_shape),
  1139. 分别对应图像np.ndarray数据、图像大小信息。
  1140. Raises:
  1141. TypeError: 形参数据类型不满足需求。
  1142. ValueError: 数据长度不匹配。
  1143. """
  1144. im = permute(im, False)
  1145. if self.mode == 'train':
  1146. if im_info is None or label_info is None:
  1147. raise TypeError(
  1148. 'Cannot do ArrangeYolov3! ' +
  1149. 'Becasuse the im_info and label_info can not be None!')
  1150. im_shape = im_info['image_shape']
  1151. if len(label_info['gt_bbox']) != len(label_info['gt_class']):
  1152. raise ValueError("gt num mismatch: bbox and class.")
  1153. if len(label_info['gt_bbox']) != len(label_info['gt_score']):
  1154. raise ValueError("gt num mismatch: bbox and score.")
  1155. gt_bbox = np.zeros((50, 4), dtype=im.dtype)
  1156. gt_class = np.zeros((50, ), dtype=np.int32)
  1157. gt_score = np.zeros((50, ), dtype=im.dtype)
  1158. gt_num = min(50, len(label_info['gt_bbox']))
  1159. if gt_num > 0:
  1160. label_info['gt_class'][:gt_num, 0] = label_info[
  1161. 'gt_class'][:gt_num, 0] - 1
  1162. if -1 not in label_info['gt_class']:
  1163. gt_bbox[:gt_num, :] = label_info['gt_bbox'][:gt_num, :]
  1164. gt_class[:gt_num] = label_info['gt_class'][:gt_num, 0]
  1165. gt_score[:gt_num] = label_info['gt_score'][:gt_num, 0]
  1166. # parse [x1, y1, x2, y2] to [x, y, w, h]
  1167. gt_bbox[:, 2:4] = gt_bbox[:, 2:4] - gt_bbox[:, :2]
  1168. gt_bbox[:, :2] = gt_bbox[:, :2] + gt_bbox[:, 2:4] / 2.
  1169. outputs = (im, gt_bbox, gt_class, gt_score, im_shape)
  1170. elif self.mode == 'eval':
  1171. if im_info is None or label_info is None:
  1172. raise TypeError(
  1173. 'Cannot do ArrangeYolov3! ' +
  1174. 'Becasuse the im_info and label_info can not be None!')
  1175. im_shape = im_info['image_shape']
  1176. if len(label_info['gt_bbox']) != len(label_info['gt_class']):
  1177. raise ValueError("gt num mismatch: bbox and class.")
  1178. im_id = im_info['im_id']
  1179. gt_bbox = np.zeros((50, 4), dtype=im.dtype)
  1180. gt_class = np.zeros((50, ), dtype=np.int32)
  1181. difficult = np.zeros((50, ), dtype=np.int32)
  1182. gt_num = min(50, len(label_info['gt_bbox']))
  1183. if gt_num > 0:
  1184. label_info['gt_class'][:gt_num, 0] = label_info[
  1185. 'gt_class'][:gt_num, 0] - 1
  1186. gt_bbox[:gt_num, :] = label_info['gt_bbox'][:gt_num, :]
  1187. gt_class[:gt_num] = label_info['gt_class'][:gt_num, 0]
  1188. difficult[:gt_num] = label_info['difficult'][:gt_num, 0]
  1189. outputs = (im, im_shape, im_id, gt_bbox, gt_class, difficult)
  1190. else:
  1191. if im_info is None:
  1192. raise TypeError('Cannot do ArrangeYolov3! ' +
  1193. 'Becasuse the im_info can not be None!')
  1194. im_shape = im_info['image_shape']
  1195. outputs = (im, im_shape)
  1196. return outputs
  1197. class ComposedRCNNTransforms(Compose):
  1198. """ RCNN模型(faster-rcnn/mask-rcnn)图像处理流程,具体如下,
  1199. 训练阶段:
  1200. 1. 随机以0.5的概率将图像水平翻转
  1201. 2. 图像归一化
  1202. 3. 图像按比例Resize,scale计算方式如下
  1203. scale = min_max_size[0] / short_size_of_image
  1204. if max_size_of_image * scale > min_max_size[1]:
  1205. scale = min_max_size[1] / max_size_of_image
  1206. 4. 将3步骤的长宽进行padding,使得长宽为32的倍数
  1207. 验证阶段:
  1208. 1. 图像归一化
  1209. 2. 图像按比例Resize,scale计算方式同上训练阶段
  1210. 3. 将2步骤的长宽进行padding,使得长宽为32的倍数
  1211. Args:
  1212. mode(str): 图像处理流程所处阶段,训练/验证/预测,分别对应'train', 'eval', 'test'
  1213. min_max_size(list): 图像在缩放时,最小边和最大边的约束条件
  1214. mean(list): 图像均值
  1215. std(list): 图像方差
  1216. random_horizontal_flip(bool): 是否以0.5的概率使用随机水平翻转增强,该仅在mode为`train`时生效,默认为True
  1217. """
  1218. def __init__(self,
  1219. mode,
  1220. min_max_size=[800, 1333],
  1221. mean=[0.485, 0.456, 0.406],
  1222. std=[0.229, 0.224, 0.225],
  1223. random_horizontal_flip=True):
  1224. if mode == 'train':
  1225. # 训练时的transforms,包含数据增强
  1226. transforms = [
  1227. Normalize(
  1228. mean=mean, std=std), ResizeByShort(
  1229. short_size=min_max_size[0], max_size=min_max_size[1]),
  1230. Padding(coarsest_stride=32)
  1231. ]
  1232. if random_horizontal_flip:
  1233. transforms.insert(0, RandomHorizontalFlip())
  1234. else:
  1235. # 验证/预测时的transforms
  1236. transforms = [
  1237. Normalize(
  1238. mean=mean, std=std), ResizeByShort(
  1239. short_size=min_max_size[0], max_size=min_max_size[1]),
  1240. Padding(coarsest_stride=32)
  1241. ]
  1242. super(ComposedRCNNTransforms, self).__init__(transforms)
  1243. class ComposedYOLOv3Transforms(Compose):
  1244. """YOLOv3模型的图像预处理流程,具体如下,
  1245. 训练阶段:
  1246. 1. 在前mixup_epoch轮迭代中,使用MixupImage策略,见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#mixupimage
  1247. 2. 对图像进行随机扰动,包括亮度,对比度,饱和度和色调
  1248. 3. 随机扩充图像,见https://paddlex.readthedocs.io/zh_CN/latest/apis/transforms/det_transforms.html#randomexpand
  1249. 4. 随机裁剪图像
  1250. 5. 将4步骤的输出图像Resize成shape参数的大小
  1251. 6. 随机0.5的概率水平翻转图像
  1252. 7. 图像归一化
  1253. 验证/预测阶段:
  1254. 1. 将图像Resize成shape参数大小
  1255. 2. 图像归一化
  1256. Args:
  1257. mode(str): 图像处理流程所处阶段,训练/验证/预测,分别对应'train', 'eval', 'test'
  1258. shape(list): 输入模型中图像的大小,输入模型的图像会被Resize成此大小
  1259. mixup_epoch(int): 模型训练过程中,前mixup_epoch会使用mixup策略, 若设为-1,则表示不使用该策略
  1260. mean(list): 图像均值
  1261. std(list): 图像方差
  1262. random_distort(bool): 数据增强方式,参数仅在mode为`train`时生效,表示是否在训练过程中随机扰动图像,默认为True
  1263. random_expand(bool): 数据增强方式,参数仅在mode为`train`时生效,表示是否在训练过程中随机扩张图像,默认为True
  1264. random_crop(bool): 数据增强方式,参数仅在mode为`train`时生效,表示是否在训练过程中随机裁剪图像,默认为True
  1265. random_horizontal_flip(bool): 数据增强方式,参数仅在mode为`train`时生效,表示是否在训练过程中随机水平翻转图像,默认为True
  1266. """
  1267. def __init__(self,
  1268. mode,
  1269. shape=[608, 608],
  1270. mixup_epoch=250,
  1271. mean=[0.485, 0.456, 0.406],
  1272. std=[0.229, 0.224, 0.225],
  1273. random_distort=True,
  1274. random_expand=True,
  1275. random_crop=True,
  1276. random_horizontal_flip=True):
  1277. width = shape
  1278. if isinstance(shape, list):
  1279. if shape[0] != shape[1]:
  1280. raise Exception(
  1281. "In YOLOv3 model, width and height should be equal")
  1282. width = shape[0]
  1283. if width % 32 != 0:
  1284. raise Exception(
  1285. "In YOLOv3 model, width and height should be multiple of 32, e.g 224、256、320...."
  1286. )
  1287. if mode == 'train':
  1288. # 训练时的transforms,包含数据增强
  1289. transforms = [
  1290. MixupImage(mixup_epoch=mixup_epoch), Resize(
  1291. target_size=width, interp='RANDOM'), Normalize(
  1292. mean=mean, std=std)
  1293. ]
  1294. if random_horizontal_flip:
  1295. transforms.insert(1, RandomHorizontalFlip())
  1296. if random_crop:
  1297. transforms.insert(1, RandomCrop())
  1298. if random_expand:
  1299. transforms.insert(1, RandomExpand())
  1300. if random_distort:
  1301. transforms.insert(1, RandomDistort())
  1302. else:
  1303. # 验证/预测时的transforms
  1304. transforms = [
  1305. Resize(
  1306. target_size=width, interp='CUBIC'), Normalize(
  1307. mean=mean, std=std)
  1308. ]
  1309. super(ComposedYOLOv3Transforms, self).__init__(transforms)
  1310. class BatchRandomShape(DetTransform):
  1311. """调整图像大小(resize)。
  1312. 对batch数据中的每张图像全部resize到random_shapes中任意一个大小。
  1313. 注意:当插值方式为“RANDOM”时,则随机选取一种插值方式进行resize。
  1314. Args:
  1315. random_shapes (list): resize大小选择列表。
  1316. 默认为[320, 352, 384, 416, 448, 480, 512, 544, 576, 608]。
  1317. interp (str): resize的插值方式,与opencv的插值方式对应,取值范围为
  1318. ['NEAREST', 'LINEAR', 'CUBIC', 'AREA', 'LANCZOS4', 'RANDOM']。默认为"RANDOM"。
  1319. Raises:
  1320. ValueError: 插值方式不在['NEAREST', 'LINEAR', 'CUBIC',
  1321. 'AREA', 'LANCZOS4', 'RANDOM']中。
  1322. """
  1323. # The interpolation mode
  1324. interp_dict = {
  1325. 'NEAREST': cv2.INTER_NEAREST,
  1326. 'LINEAR': cv2.INTER_LINEAR,
  1327. 'CUBIC': cv2.INTER_CUBIC,
  1328. 'AREA': cv2.INTER_AREA,
  1329. 'LANCZOS4': cv2.INTER_LANCZOS4
  1330. }
  1331. def __init__(
  1332. self,
  1333. random_shapes=[320, 352, 384, 416, 448, 480, 512, 544, 576, 608],
  1334. interp='RANDOM'):
  1335. if not (interp == "RANDOM" or interp in self.interp_dict):
  1336. raise ValueError("interp should be one of {}".format(
  1337. self.interp_dict.keys()))
  1338. self.random_shapes = random_shapes
  1339. self.interp = interp
  1340. def __call__(self, batch_data):
  1341. """
  1342. Args:
  1343. batch_data (list): 由与图像相关的各种信息组成的batch数据。
  1344. Returns:
  1345. list: 由与图像相关的各种信息组成的batch数据。
  1346. """
  1347. shape = np.random.choice(self.random_shapes)
  1348. if self.interp == "RANDOM":
  1349. interp = random.choice(list(self.interp_dict.keys()))
  1350. else:
  1351. interp = self.interp
  1352. for data_id, data in enumerate(batch_data):
  1353. data_list = list(data)
  1354. im = data_list[0]
  1355. im = np.swapaxes(im, 1, 0)
  1356. im = np.swapaxes(im, 1, 2)
  1357. im = resize(im, shape, self.interp_dict[interp])
  1358. im = np.swapaxes(im, 1, 2)
  1359. im = np.swapaxes(im, 1, 0)
  1360. data_list[0] = im
  1361. batch_data[data_id] = tuple(data_list)
  1362. return batch_data
  1363. class GenerateYoloTarget(object):
  1364. """生成YOLOv3的ground truth(真实标注框)在不同特征层的位置转换信息。
  1365. 该transform只在YOLOv3计算细粒度loss时使用。
  1366. Args:
  1367. anchors (list|tuple): anchor框的宽度和高度。
  1368. anchor_masks (list|tuple): 在计算损失时,使用anchor的mask索引。
  1369. num_classes (int): 类别数。默认为80。
  1370. iou_thresh (float): iou阈值,当anchor和真实标注框的iou大于该阈值时,计入target。默认为1.0。
  1371. """
  1372. def __init__(self,
  1373. anchors,
  1374. anchor_masks,
  1375. downsample_ratios,
  1376. num_classes=80,
  1377. iou_thresh=1.):
  1378. super(GenerateYoloTarget, self).__init__()
  1379. self.anchors = anchors
  1380. self.anchor_masks = anchor_masks
  1381. self.downsample_ratios = downsample_ratios
  1382. self.num_classes = num_classes
  1383. self.iou_thresh = iou_thresh
  1384. def __call__(self, batch_data):
  1385. """
  1386. Args:
  1387. batch_data (list): 由与图像相关的各种信息组成的batch数据。
  1388. Returns:
  1389. list: 由与图像相关的各种信息组成的batch数据。
  1390. 其中,每个数据新添加的字段为:
  1391. - target0 (np.ndarray): YOLOv3的ground truth在特征层0的位置转换信息,
  1392. 形状为(特征层0的anchor数量, 6+类别数, 特征层0的h, 特征层0的w)。
  1393. - target1 (np.ndarray): YOLOv3的ground truth在特征层1的位置转换信息,
  1394. 形状为(特征层1的anchor数量, 6+类别数, 特征层1的h, 特征层1的w)。
  1395. - ...
  1396. -targetn (np.ndarray): YOLOv3的ground truth在特征层n的位置转换信息,
  1397. 形状为(特征层n的anchor数量, 6+类别数, 特征层n的h, 特征层n的w)。
  1398. n的是大小由anchor_masks的长度决定。
  1399. """
  1400. im = batch_data[0][0]
  1401. h = im.shape[1]
  1402. w = im.shape[2]
  1403. an_hw = np.array(self.anchors) / np.array([[w, h]])
  1404. for data_id, data in enumerate(batch_data):
  1405. gt_bbox = data[1]
  1406. gt_class = data[2]
  1407. gt_score = data[3]
  1408. im_shape = data[4]
  1409. origin_h = float(im_shape[0])
  1410. origin_w = float(im_shape[1])
  1411. data_list = list(data)
  1412. for i, (
  1413. mask, downsample_ratio
  1414. ) in enumerate(zip(self.anchor_masks, self.downsample_ratios)):
  1415. grid_h = int(h / downsample_ratio)
  1416. grid_w = int(w / downsample_ratio)
  1417. target = np.zeros(
  1418. (len(mask), 6 + self.num_classes, grid_h, grid_w),
  1419. dtype=np.float32)
  1420. for b in range(gt_bbox.shape[0]):
  1421. gx = gt_bbox[b, 0] / float(origin_w)
  1422. gy = gt_bbox[b, 1] / float(origin_h)
  1423. gw = gt_bbox[b, 2] / float(origin_w)
  1424. gh = gt_bbox[b, 3] / float(origin_h)
  1425. cls = gt_class[b]
  1426. score = gt_score[b]
  1427. if gw <= 0. or gh <= 0. or score <= 0.:
  1428. continue
  1429. # find best match anchor index
  1430. best_iou = 0.
  1431. best_idx = -1
  1432. for an_idx in range(an_hw.shape[0]):
  1433. iou = jaccard_overlap(
  1434. [0., 0., gw, gh],
  1435. [0., 0., an_hw[an_idx, 0], an_hw[an_idx, 1]])
  1436. if iou > best_iou:
  1437. best_iou = iou
  1438. best_idx = an_idx
  1439. gi = int(gx * grid_w)
  1440. gj = int(gy * grid_h)
  1441. # gtbox should be regresed in this layes if best match
  1442. # anchor index in anchor mask of this layer
  1443. if best_idx in mask:
  1444. best_n = mask.index(best_idx)
  1445. # x, y, w, h, scale
  1446. target[best_n, 0, gj, gi] = gx * grid_w - gi
  1447. target[best_n, 1, gj, gi] = gy * grid_h - gj
  1448. target[best_n, 2, gj, gi] = np.log(
  1449. gw * w / self.anchors[best_idx][0])
  1450. target[best_n, 3, gj, gi] = np.log(
  1451. gh * h / self.anchors[best_idx][1])
  1452. target[best_n, 4, gj, gi] = 2.0 - gw * gh
  1453. # objectness record gt_score
  1454. target[best_n, 5, gj, gi] = score
  1455. # classification
  1456. target[best_n, 6 + cls, gj, gi] = 1.
  1457. # For non-matched anchors, calculate the target if the iou
  1458. # between anchor and gt is larger than iou_thresh
  1459. if self.iou_thresh < 1:
  1460. for idx, mask_i in enumerate(mask):
  1461. if mask_i == best_idx: continue
  1462. iou = jaccard_overlap(
  1463. [0., 0., gw, gh],
  1464. [0., 0., an_hw[mask_i, 0], an_hw[mask_i, 1]])
  1465. if iou > self.iou_thresh:
  1466. # x, y, w, h, scale
  1467. target[idx, 0, gj, gi] = gx * grid_w - gi
  1468. target[idx, 1, gj, gi] = gy * grid_h - gj
  1469. target[idx, 2, gj, gi] = np.log(
  1470. gw * w / self.anchors[mask_i][0])
  1471. target[idx, 3, gj, gi] = np.log(
  1472. gh * h / self.anchors[mask_i][1])
  1473. target[idx, 4, gj, gi] = 2.0 - gw * gh
  1474. # objectness record gt_score
  1475. target[idx, 5, gj, gi] = score
  1476. # classification
  1477. target[idx, 6 + cls, gj, gi] = 1.
  1478. data_list.append(target)
  1479. batch_data[data_id] = tuple(data_list)
  1480. return batch_data