k20hcmus/FishEye8K
3
1import math2import random3 4import cv25import numpy as np6import torch7import torchvision.transforms as T8import torchvision.transforms.functional as TF9 10from utils.general import LOGGER, check_version, colorstr, resample_segments, segment2box, xywhn2xyxy11from utils.metrics import bbox_ioa12 13IMAGENET_MEAN = 0.485, 0.456, 0.406 # RGB mean14IMAGENET_STD = 0.229, 0.224, 0.225 # RGB standard deviation15 16 17class Albumentations:18 # YOLOv5 Albumentations class (optional, only used if package is installed)19 def __init__(self, size=640):20 self.transform = None21 prefix = colorstr('albumentations: ')22 try:23 import albumentations as A24 check_version(A.__version__, '1.0.3', hard=True) # version requirement25 26 T = [27 A.RandomResizedCrop(height=size, width=size, scale=(0.8, 1.0), ratio=(0.9, 1.11), p=0.0),28 A.Blur(p=0.01),29 A.MedianBlur(p=0.01),30 A.ToGray(p=0.01),31 A.CLAHE(p=0.01),32 A.RandomBrightnessContrast(p=0.0),33 A.RandomGamma(p=0.0),34 A.ImageCompression(quality_lower=75, p=0.0)] # transforms35 self.transform = A.Compose(T, bbox_params=A.BboxParams(format='yolo', label_fields=['class_labels']))36 37 LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))38 except ImportError: # package not installed, skip39 pass40 except Exception as e:41 LOGGER.info(f'{prefix}{e}')42 43 def __call__(self, im, labels, p=1.0):44 if self.transform and random.random() < p:45 new = self.transform(image=im, bboxes=labels[:, 1:], class_labels=labels[:, 0]) # transformed46 im, labels = new['image'], np.array([[c, *b] for c, b in zip(new['class_labels'], new['bboxes'])])47 return im, labels48 49 50def normalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD, inplace=False):51 # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = (x - mean) / std52 return TF.normalize(x, mean, std, inplace=inplace)53 54 55def denormalize(x, mean=IMAGENET_MEAN, std=IMAGENET_STD):56 # Denormalize RGB images x per ImageNet stats in BCHW format, i.e. = x * std + mean57 for i in range(3):58 x[:, i] = x[:, i] * std[i] + mean[i]59 return x60 61 62def augment_hsv(im, hgain=0.5, sgain=0.5, vgain=0.5):63 # HSV color-space augmentation64 if hgain or sgain or vgain:65 r = np.random.uniform(-1, 1, 3) * [hgain, sgain, vgain] + 1 # random gains66 hue, sat, val = cv2.split(cv2.cvtColor(im, cv2.COLOR_BGR2HSV))67 dtype = im.dtype # uint868 69 x = np.arange(0, 256, dtype=r.dtype)70 lut_hue = ((x * r[0]) % 180).astype(dtype)71 lut_sat = np.clip(x * r[1], 0, 255).astype(dtype)72 lut_val = np.clip(x * r[2], 0, 255).astype(dtype)73 74 im_hsv = cv2.merge((cv2.LUT(hue, lut_hue), cv2.LUT(sat, lut_sat), cv2.LUT(val, lut_val)))75 cv2.cvtColor(im_hsv, cv2.COLOR_HSV2BGR, dst=im) # no return needed76 77 78def hist_equalize(im, clahe=True, bgr=False):79 # Equalize histogram on BGR image 'im' with im.shape(n,m,3) and range 0-25580 yuv = cv2.cvtColor(im, cv2.COLOR_BGR2YUV if bgr else cv2.COLOR_RGB2YUV)81 if clahe:82 c = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))83 yuv[:, :, 0] = c.apply(yuv[:, :, 0])84 else:85 yuv[:, :, 0] = cv2.equalizeHist(yuv[:, :, 0]) # equalize Y channel histogram86 return cv2.cvtColor(yuv, cv2.COLOR_YUV2BGR if bgr else cv2.COLOR_YUV2RGB) # convert YUV image to RGB87 88 89def replicate(im, labels):90 # Replicate labels91 h, w = im.shape[:2]92 boxes = labels[:, 1:].astype(int)93 x1, y1, x2, y2 = boxes.T94 s = ((x2 - x1) + (y2 - y1)) / 2 # side length (pixels)95 for i in s.argsort()[:round(s.size * 0.5)]: # smallest indices96 x1b, y1b, x2b, y2b = boxes[i]97 bh, bw = y2b - y1b, x2b - x1b98 yc, xc = int(random.uniform(0, h - bh)), int(random.uniform(0, w - bw)) # offset x, y99 x1a, y1a, x2a, y2a = [xc, yc, xc + bw, yc + bh]100 im[y1a:y2a, x1a:x2a] = im[y1b:y2b, x1b:x2b] # im4[ymin:ymax, xmin:xmax]101 labels = np.append(labels, [[labels[i, 0], x1a, y1a, x2a, y2a]], axis=0)102 103 return im, labels104 105 106def letterbox(im, new_shape=(640, 640), color=(114, 114, 114), auto=True, scaleFill=False, scaleup=True, stride=32):107 # Resize and pad image while meeting stride-multiple constraints108 shape = im.shape[:2] # current shape [height, width]109 if isinstance(new_shape, int):110 new_shape = (new_shape, new_shape)111 112 # Scale ratio (new / old)113 r = min(new_shape[0] / shape[0], new_shape[1] / shape[1])114 if not scaleup: # only scale down, do not scale up (for better val mAP)115 r = min(r, 1.0)116 117 # Compute padding118 ratio = r, r # width, height ratios119 new_unpad = int(round(shape[1] * r)), int(round(shape[0] * r))120 dw, dh = new_shape[1] - new_unpad[0], new_shape[0] - new_unpad[1] # wh padding121 if auto: # minimum rectangle122 dw, dh = np.mod(dw, stride), np.mod(dh, stride) # wh padding123 elif scaleFill: # stretch124 dw, dh = 0.0, 0.0125 new_unpad = (new_shape[1], new_shape[0])126 ratio = new_shape[1] / shape[1], new_shape[0] / shape[0] # width, height ratios127 128 dw /= 2 # divide padding into 2 sides129 dh /= 2130 131 if shape[::-1] != new_unpad: # resize132 im = cv2.resize(im, new_unpad, interpolation=cv2.INTER_LINEAR)133 top, bottom = int(round(dh - 0.1)), int(round(dh + 0.1))134 left, right = int(round(dw - 0.1)), int(round(dw + 0.1))135 im = cv2.copyMakeBorder(im, top, bottom, left, right, cv2.BORDER_CONSTANT, value=color) # add border136 return im, ratio, (dw, dh)137 138 139def random_perspective(im,140 targets=(),141 segments=(),142 degrees=10,143 translate=.1,144 scale=.1,145 shear=10,146 perspective=0.0,147 border=(0, 0)):148 # torchvision.transforms.RandomAffine(degrees=(-10, 10), translate=(0.1, 0.1), scale=(0.9, 1.1), shear=(-10, 10))149 # targets = [cls, xyxy]150 151 height = im.shape[0] + border[0] * 2 # shape(h,w,c)152 width = im.shape[1] + border[1] * 2153 154 # Center155 C = np.eye(3)156 C[0, 2] = -im.shape[1] / 2 # x translation (pixels)157 C[1, 2] = -im.shape[0] / 2 # y translation (pixels)158 159 # Perspective160 P = np.eye(3)161 P[2, 0] = random.uniform(-perspective, perspective) # x perspective (about y)162 P[2, 1] = random.uniform(-perspective, perspective) # y perspective (about x)163 164 # Rotation and Scale165 R = np.eye(3)166 a = random.uniform(-degrees, degrees)167 # a += random.choice([-180, -90, 0, 90]) # add 90deg rotations to small rotations168 s = random.uniform(1 - scale, 1 + scale)169 # s = 2 ** random.uniform(-scale, scale)170 R[:2] = cv2.getRotationMatrix2D(angle=a, center=(0, 0), scale=s)171 172 # Shear173 S = np.eye(3)174 S[0, 1] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # x shear (deg)175 S[1, 0] = math.tan(random.uniform(-shear, shear) * math.pi / 180) # y shear (deg)176 177 # Translation178 T = np.eye(3)179 T[0, 2] = random.uniform(0.5 - translate, 0.5 + translate) * width # x translation (pixels)180 T[1, 2] = random.uniform(0.5 - translate, 0.5 + translate) * height # y translation (pixels)181 182 # Combined rotation matrix183 M = T @ S @ R @ P @ C # order of operations (right to left) is IMPORTANT184 if (border[0] != 0) or (border[1] != 0) or (M != np.eye(3)).any(): # image changed185 if perspective:186 im = cv2.warpPerspective(im, M, dsize=(width, height), borderValue=(114, 114, 114))187 else: # affine188 im = cv2.warpAffine(im, M[:2], dsize=(width, height), borderValue=(114, 114, 114))189 190 # Visualize191 # import matplotlib.pyplot as plt192 # ax = plt.subplots(1, 2, figsize=(12, 6))[1].ravel()193 # ax[0].imshow(im[:, :, ::-1]) # base194 # ax[1].imshow(im2[:, :, ::-1]) # warped195 196 # Transform label coordinates197 n = len(targets)198 if n:199 use_segments = any(x.any() for x in segments)200 new = np.zeros((n, 4))201 if use_segments: # warp segments202 segments = resample_segments(segments) # upsample203 for i, segment in enumerate(segments):204 xy = np.ones((len(segment), 3))205 xy[:, :2] = segment206 xy = xy @ M.T # transform207 xy = xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2] # perspective rescale or affine208 209 # clip210 new[i] = segment2box(xy, width, height)211 212 else: # warp boxes213 xy = np.ones((n * 4, 3))214 xy[:, :2] = targets[:, [1, 2, 3, 4, 1, 4, 3, 2]].reshape(n * 4, 2) # x1y1, x2y2, x1y2, x2y1215 xy = xy @ M.T # transform216 xy = (xy[:, :2] / xy[:, 2:3] if perspective else xy[:, :2]).reshape(n, 8) # perspective rescale or affine217 218 # create new boxes219 x = xy[:, [0, 2, 4, 6]]220 y = xy[:, [1, 3, 5, 7]]221 new = np.concatenate((x.min(1), y.min(1), x.max(1), y.max(1))).reshape(4, n).T222 223 # clip224 new[:, [0, 2]] = new[:, [0, 2]].clip(0, width)225 new[:, [1, 3]] = new[:, [1, 3]].clip(0, height)226 227 # filter candidates228 i = box_candidates(box1=targets[:, 1:5].T * s, box2=new.T, area_thr=0.01 if use_segments else 0.10)229 targets = targets[i]230 targets[:, 1:5] = new[i]231 232 return im, targets233 234 235def copy_paste(im, labels, segments, p=0.5):236 # Implement Copy-Paste augmentation https://arxiv.org/abs/2012.07177, labels as nx5 np.array(cls, xyxy)237 n = len(segments)238 if p and n:239 h, w, c = im.shape # height, width, channels240 im_new = np.zeros(im.shape, np.uint8)241 242 # calculate ioa first then select indexes randomly243 boxes = np.stack([w - labels[:, 3], labels[:, 2], w - labels[:, 1], labels[:, 4]], axis=-1) # (n, 4)244 ioa = bbox_ioa(boxes, labels[:, 1:5]) # intersection over area245 indexes = np.nonzero((ioa < 0.30).all(1))[0] # (N, )246 n = len(indexes)247 for j in random.sample(list(indexes), k=round(p * n)):248 l, box, s = labels[j], boxes[j], segments[j]249 labels = np.concatenate((labels, [[l[0], *box]]), 0)250 segments.append(np.concatenate((w - s[:, 0:1], s[:, 1:2]), 1))251 cv2.drawContours(im_new, [segments[j].astype(np.int32)], -1, (1, 1, 1), cv2.FILLED)252 253 result = cv2.flip(im, 1) # augment segments (flip left-right)254 i = cv2.flip(im_new, 1).astype(bool)255 im[i] = result[i] # cv2.imwrite('debug.jpg', im) # debug256 257 return im, labels, segments258 259 260def cutout(im, labels, p=0.5):261 # Applies image cutout augmentation https://arxiv.org/abs/1708.04552262 if random.random() < p:263 h, w = im.shape[:2]264 scales = [0.5] * 1 + [0.25] * 2 + [0.125] * 4 + [0.0625] * 8 + [0.03125] * 16 # image size fraction265 for s in scales:266 mask_h = random.randint(1, int(h * s)) # create random masks267 mask_w = random.randint(1, int(w * s))268 269 # box270 xmin = max(0, random.randint(0, w) - mask_w // 2)271 ymin = max(0, random.randint(0, h) - mask_h // 2)272 xmax = min(w, xmin + mask_w)273 ymax = min(h, ymin + mask_h)274 275 # apply random color mask276 im[ymin:ymax, xmin:xmax] = [random.randint(64, 191) for _ in range(3)]277 278 # return unobscured labels279 if len(labels) and s > 0.03:280 box = np.array([[xmin, ymin, xmax, ymax]], dtype=np.float32)281 ioa = bbox_ioa(box, xywhn2xyxy(labels[:, 1:5], w, h))[0] # intersection over area282 labels = labels[ioa < 0.60] # remove >60% obscured labels283 284 return labels285 286 287def mixup(im, labels, im2, labels2):288 # Applies MixUp augmentation https://arxiv.org/pdf/1710.09412.pdf289 r = np.random.beta(32.0, 32.0) # mixup ratio, alpha=beta=32.0290 im = (im * r + im2 * (1 - r)).astype(np.uint8)291 labels = np.concatenate((labels, labels2), 0)292 return im, labels293 294 295def box_candidates(box1, box2, wh_thr=2, ar_thr=100, area_thr=0.1, eps=1e-16): # box1(4,n), box2(4,n)296 # Compute candidate boxes: box1 before augment, box2 after augment, wh_thr (pixels), aspect_ratio_thr, area_ratio297 w1, h1 = box1[2] - box1[0], box1[3] - box1[1]298 w2, h2 = box2[2] - box2[0], box2[3] - box2[1]299 ar = np.maximum(w2 / (h2 + eps), h2 / (w2 + eps)) # aspect ratio300 return (w2 > wh_thr) & (h2 > wh_thr) & (w2 * h2 / (w1 * h1 + eps) > area_thr) & (ar < ar_thr) # candidates301 302 303def classify_albumentations(304 augment=True,305 size=224,306 scale=(0.08, 1.0),307 ratio=(0.75, 1.0 / 0.75), # 0.75, 1.33308 hflip=0.5,309 vflip=0.0,310 jitter=0.4,311 mean=IMAGENET_MEAN,312 std=IMAGENET_STD,313 auto_aug=False):314 # YOLOv5 classification Albumentations (optional, only used if package is installed)315 prefix = colorstr('albumentations: ')316 try:317 import albumentations as A318 from albumentations.pytorch import ToTensorV2319 check_version(A.__version__, '1.0.3', hard=True) # version requirement320 if augment: # Resize and crop321 T = [A.RandomResizedCrop(height=size, width=size, scale=scale, ratio=ratio)]322 if auto_aug:323 # TODO: implement AugMix, AutoAug & RandAug in albumentation324 LOGGER.info(f'{prefix}auto augmentations are currently not supported')325 else:326 if hflip > 0:327 T += [A.HorizontalFlip(p=hflip)]328 if vflip > 0:329 T += [A.VerticalFlip(p=vflip)]330 if jitter > 0:331 color_jitter = (float(jitter),) * 3 # repeat value for brightness, contrast, satuaration, 0 hue332 T += [A.ColorJitter(*color_jitter, 0)]333 else: # Use fixed crop for eval set (reproducibility)334 T = [A.SmallestMaxSize(max_size=size), A.CenterCrop(height=size, width=size)]335 T += [A.Normalize(mean=mean, std=std), ToTensorV2()] # Normalize and convert to Tensor336 LOGGER.info(prefix + ', '.join(f'{x}'.replace('always_apply=False, ', '') for x in T if x.p))337 return A.Compose(T)338 339 except ImportError: # package not installed, skip340 LOGGER.warning(f'{prefix}⚠️ not found, install with `pip install albumentations` (recommended)')341 except Exception as e:342 LOGGER.info(f'{prefix}{e}')343 344 345def classify_transforms(size=224):346 # Transforms to apply if albumentations not installed347 assert isinstance(size, int), f'ERROR: classify_transforms size {size} must be integer, not (list, tuple)'348 # T.Compose([T.ToTensor(), T.Resize(size), T.CenterCrop(size), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])349 return T.Compose([CenterCrop(size), ToTensor(), T.Normalize(IMAGENET_MEAN, IMAGENET_STD)])350 351 352class LetterBox:353 # YOLOv5 LetterBox class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])354 def __init__(self, size=(640, 640), auto=False, stride=32):355 super().__init__()356 self.h, self.w = (size, size) if isinstance(size, int) else size357 self.auto = auto # pass max size integer, automatically solve for short side using stride358 self.stride = stride # used with auto359 360 def __call__(self, im): # im = np.array HWC361 imh, imw = im.shape[:2]362 r = min(self.h / imh, self.w / imw) # ratio of new/old363 h, w = round(imh * r), round(imw * r) # resized image364 hs, ws = (math.ceil(x / self.stride) * self.stride for x in (h, w)) if self.auto else self.h, self.w365 top, left = round((hs - h) / 2 - 0.1), round((ws - w) / 2 - 0.1)366 im_out = np.full((self.h, self.w, 3), 114, dtype=im.dtype)367 im_out[top:top + h, left:left + w] = cv2.resize(im, (w, h), interpolation=cv2.INTER_LINEAR)368 return im_out369 370 371class CenterCrop:372 # YOLOv5 CenterCrop class for image preprocessing, i.e. T.Compose([CenterCrop(size), ToTensor()])373 def __init__(self, size=640):374 super().__init__()375 self.h, self.w = (size, size) if isinstance(size, int) else size376 377 def __call__(self, im): # im = np.array HWC378 imh, imw = im.shape[:2]379 m = min(imh, imw) # min dimension380 top, left = (imh - m) // 2, (imw - m) // 2381 return cv2.resize(im[top:top + m, left:left + m], (self.w, self.h), interpolation=cv2.INTER_LINEAR)382 383 384class ToTensor:385 # YOLOv5 ToTensor class for image preprocessing, i.e. T.Compose([LetterBox(size), ToTensor()])386 def __init__(self, half=False):387 super().__init__()388 self.half = half389 390 def __call__(self, im): # im = np.array HWC in BGR order391 im = np.ascontiguousarray(im.transpose((2, 0, 1))[::-1]) # HWC to CHW -> BGR to RGB -> contiguous392 im = torch.from_numpy(im) # to torch393 im = im.half() if self.half else im.float() # uint8 to fp16/32394 im /= 255.0 # 0-255 to 0.0-1.0395 return im396