def _prepare_blobs(im, pixel_means, target_size, max_size):
""" Reference: blob.prep_im_for_blob() """
im = im.astype(np.float32, copy=False)
im -= pixel_means
if cfg.SOBEL:
###### add by Xiaohang ########
im /= np.array([[[57.375, 57.12, 58.395]]])
im = sobel(im)
###############################
im_shape = im.shape
im_size_min = np.min(im_shape[0:2])
im_size_max = np.max(im_shape[0:2])
im_scale = float(target_size) / float(im_size_min)
if np.round(im_scale * im_size_max) > max_size:
im_scale = float(max_size) / float(im_size_max)
im = cv2.resize(im,
None,
None,
fx=im_scale,
fy=im_scale,
interpolation=cv2.INTER_LINEAR)
# Reuse code in blob_utils and fit FPN
blob = blob_utils.im_list_to_blob([im])
blobs = {}
blobs["data"] = blob
blobs["im_info"] = np.array([[blob.shape[2], blob.shape[3], im_scale]],
dtype=np.float32)
return blobs
def _prepare_blobs(
im,
pixel_means,
target_size,
max_size,
):
''' Reference: blob.prep_im_for_blob() '''
im = im.astype(np.float32, copy=False)
im -= pixel_means
im_shape = im.shape
im_size_min = np.min(im_shape[0:2])
im_size_max = np.max(im_shape[0:2])
im_scale = float(target_size) / float(im_size_min)
if np.round(im_scale * im_size_max) > max_size:
im_scale = float(max_size) / float(im_size_max)
im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale,
interpolation=cv2.INTER_LINEAR)
# Reuse code in blob_utils and fit FPN
blob = blob_utils.im_list_to_blob([im])
blobs = {}
blobs['data'] = blob
blobs['im_info'] = np.array(
[[blob.shape[2], blob.shape[3], im_scale]],
dtype=np.float32
)
return blobs
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES), size=num_images
)
processed_ims = []
im_scales = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE
)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales
def _prepare_blobs(
im,
pixel_means,
target_size,
max_size,
):
''' Reference: blob.prep_im_for_blob() '''
im = im.astype(np.float32, copy=False)
im -= pixel_means
im_shape = im.shape
im_size_min = np.min(im_shape[0:2])
im_size_max = np.max(im_shape[0:2])
im_scale = float(target_size) / float(im_size_min)
if np.round(im_scale * im_size_max) > max_size:
im_scale = float(max_size) / float(im_size_max)
im = cv2.resize(im, None, None, fx=im_scale, fy=im_scale,
interpolation=cv2.INTER_LINEAR)
# Reuse code in blob_utils and fit FPN
blob = blob_utils.im_list_to_blob([im])
blobs = {}
blobs['data'] = blob
blobs['im_info'] = np.array(
[[blob.shape[2], blob.shape[3], im_scale]],
dtype=np.float32
)
return blobs
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES), size=num_images
)
processed_ims = []
im_scales = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
logger.info(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE
)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb) # roidb是一个list数据,每一个元素是dict,获取roidb的数据个数,即样本数
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images) # 生成每一张对应的scale尺度的list
processed_ims = []
im_scales = []
Gauss_kernel = 0
SigmaX = 0
alpha = 1
beta = 0
for i in range(num_images):
im = cv2.imread(roidb[i]['image']) # 读取对应的数据图片
if cfg.TRAIN.AUGUMENT:
im_tmp = cv2.imread(roidb[i]['image'])
random_flag = random.randint(0, 1) # 随机进行高斯模糊
#print(random_flag)
if random_flag:
Gauss_kernel = 7 #np.random.randint(5, 7)
SigmaX = 5.5 #np.random.randint(0, 70) /10.0
im = randomGaussian(im_tmp, Gauss_kernel, SigmaX)
alpha = np.random.randint(0, 15) / 10.0
beta = np.random.randint(0, 15)
im = Contrast_and_Brightness(alpha, beta, im)
#cv2.imwrite('test.jpg',im)
else:
im = im_tmp.copy()
#print(random_flag,alpha,beta)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']: # 如果flipped为真 则翻转数据
im = im[:, ::-1, :] # 第二个,前的-1表示对x轴的数据反向读取,即对图进行x轴的翻转
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size,
cfg.TRAIN.MAX_SIZE) # 会做一波预处理,resize为宽为600的图
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
angle = np.random.uniform(cfg.TRAIN.ROTATION_ANGLES[0],
cfg.TRAIN.ROTATION_ANGLES[1])
shear = np.random.uniform(cfg.TRAIN.SHEAR_INT[0], cfg.TRAIN.SHEAR_INT[1])
# zoom = np.random.uniform(cfg.TRAIN.ZOOM[0], cfg.TRAIN.ZOOM[1])
channel_shift = np.random.uniform(cfg.TRAIN.CHANNEL_SHIFT[0],
cfg.TRAIN.CHANNEL_SHIFT[1])
processed_ims = []
im_transformation_matrices = []
im_shapes = []
zooms = []
im_scales = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['h_flipped'] and roidb[i]['v_flipped']:
im = im[::-1, ::-1, :]
elif roidb[i]['v_flipped']:
im = im[::-1, :, :]
elif roidb[i]['h_flipped']:
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, transformation_matrix, im_shape, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE, angle, shear,
channel_shift)
im_transformation_matrices.append(transformation_matrix)
im_shapes.append(im_shape)
# zooms.append(zoom)
processed_ims.append(im)
im_scales.append(im_scale)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_transformation_matrices, im_shapes, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
for i in range(num_images):
roi = roidb[i]
im = cv2.imread(roi['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roi['image'])
if roi['flipped']:
im = im[:, ::-1, :]
# added by Jerome
h, w, _ = im.shape
roi['width'] = w
roi['height'] = h
boxes = roi['boxes']
# will be only true for live_targets:
if cfg.TRAIN.DOMAIN_ADAPTATION:
if len(boxes) == 1 and (boxes[0] == np.array(
[0., 0., 0., 0.], dtype=np.float32)).all():
roi['boxes'] = np.array(
[[w * .2, h * .2, w * .8, h * .8]], dtype=np.float32
) # to make sure enough (one or more) anchors are retained
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
im_crops = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
# data augmentation
# prefix = uuid.uuid1()
# reid_roi_data.save_image(im, '{}_b'.format(prefix))
im, im_crop = reid_roi_data.random_crop(im)
im, im_crop2 = reid_roi_data.horizontal_crop(im)
im_crop[2] = im_crop[2] - (im_crop[2] - im_crop[0] - im_crop2[2])
im = reid_roi_data.hsv_jitter(im)
im = reid_roi_data.gaussian_blur(im)
im = reid_roi_data.random_erasing(im)
# reid_roi_data.save_image(im, '{}_a'.format(prefix))
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
im_crops.append(im_crop)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales, im_crops
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
#print("qwryyudfggjhgkjgk{}sdfgfdsg".format(roidb))
# print("minibatch.py 118 此轮roidb的长度为:",len(roidb))
# 获得数据的数量
num_images = len(roidb)
# 在这批数据中最XXXX的采样随机比例 (600,)这个的len()只是1 np.random.randint即在第一个参数和'high'值-1,的整数间生成list,list元素数量是roidb的数量
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES),
size=num_images) # 因为cfg.TRAIN.SCALES为(600,), 所以生成的是一个全为0的list
processed_ims = [] # 图片数据的list
im_scales = [] # 尺度比例list
for i in range(num_images):
#print("minibatch.py 109 @@@@@@@ roidb的boxes{}".format(roidb[i]['boxes']))
try:
im = cv2.imread(roidb[i]['image']) # 读取roidb的图像数据
#print("---这是此次的for循环的第{}/{}轮,图像为{}".format(i,num_images,roidb[i]['image']))
except Exception as e:
print("Exception:", e)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image']) # 如果图像不存在则assert
if roidb[i]['flipped']: # 如果图的flipped为翻转,则翻转
im = im[:, ::-1, :] # 对图像的宽做倒序 即x轴flipped
# s_t1 = time.time()
target_size = cfg.TRAIN.SCALES[scale_inds[i]] # target_size是600
im, im_scale = blob_utils.prep_im_for_blob( # im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale) # 加入到im_scales的list中
processed_ims.append(im) # 加入到图像数据list中
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
#e_t1 = time.time()
#t_t1 = e_t1 - s_t1
#print("时间_最终：{}".format(t_t1))
return blob, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
if roidb[i]['cropped'] != []:
width = im.shape[1]
height = im.shape[0]
crop_size = cfg.TRAIN.CROPPED_SIZE
crop_index = roidb[i]['cropped'].index(1)
xmin = int(height / crop_size * int(round(crop_index / 2)))
xmax = int(xmin + (height / crop_size))
ymin = int(width / crop_size * int(round(crop_index / 2)))
ymax = int(ymin + (width / crop_size))
im = im[xmin:xmax, ymin:ymax, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales
def _get_img_blob_mul(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
all_roidb = []
for i in range(num_images):
img_path = roidb[i]['image']
img = cv2.imread(img_path)
h, w = img.shape[:2]
# 如果图的flipped为翻转,则翻转
if roidb[i]['flipped']:
img = img[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
# print("copy图像时间：{}".format(t_t_))
coor, new_boxes_crop, img_crop_w, img_crop_h, index_list = crop(
h, w, roidb[i], WIDTH_REGION, HEIGHT_REGION)
# 深拷贝以避免改变原始图像数据
img_ = copy.deepcopy(img_[coor[0]:coor[1], coor[2]:coor[3]])
cv2.imwrite("crop_image.jpg", img_)
if not new_boxes_crop == []:
new_roidb = creat_new_roidb(roidb[i], img_crop_w, img_crop_h,
new_boxes_crop, index_list)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
time_start = time.time()
if MASKRCNN_SWITCH:
segms = []
seg_areas = []
#print("roidb:---------------------",new_roidb )
if len(new_roidb['segms']) > 0:
for m_s, roi_s in enumerate(new_roidb['segms']):
roi_s = np.array(roi_s[0][:-4]).reshape(1, -1, 2)
img_se = np.zeros(
(roidb[i]['height'], roidb[i]['width'], 3))
img_se_ = img_se.copy()[coor[0]:coor[1],
coor[2]:coor[3]]
imag = cv2.drawContours(img_se, [roi_s], -1,
(255, 255, 255), -1)
#cv2.imwrite("mask_raw.jpg", imag)
imag = imag[coor[0]:coor[1], coor[2]:coor[3]]
cv2.imwrite("mask_single.jpg", imag)
imag = cv2.imread("mask_single.jpg")
gray = cv2.cvtColor(imag, cv2.COLOR_BGR2GRAY)
# 一定要做二值化,否则出来的是散点
ret, binary = cv2.threshold(gray, 127, 255,
cv2.THRESH_BINARY)
_, contours, hierarchy = cv2.findContours(
binary, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
#img_fill = cv2.drawContours(imag, contours,-1,(0,255,0),-1)
#cv2.imwrite("mask_single_fill.jpg", img_fill)
if len(contours) > 0:
#print("------------------contours:-----------------",contours)
for kk, con_point in enumerate(contours):
if con_point[0][0][0] == 0:
con_point[0][0][0] = 1
elif con_point[0][0][0] == img_crop_w - 1:
con_point[0][0][0] = img_crop_w - 2
else:
pass
if con_point[0][0][1] == 0:
con_point[0][0][1] = 1
elif con_point[0][0][1] == img_crop_h - 1:
con_point[0][0][1] = img_crop_h - 2
else:
pass
roi_s = list(np.array(contours).reshape(1, -1))
area = float(int(cv2.contourArea(contours[0])))
segms.append(list(roi_s))
seg_areas.append(area)
segms_ = []
# 验证mask框-----
# for rk, roi_k in enumerate(segms):
# roi_k = np.array(roi_k).reshape(-1,1,2)
# segms_.append(roi_k)
# #print("!!!!!points!!!!!!!",segms_)
# imag_ = cv2.drawContours(img_se_, segms_ ,-1,(255,255,255),-1)
# cv2.imwrite("maskrcnn_crop.jpg", imag_)
# ---------------------------------------------------------------------------------
new_roidb['segms'] = list(segms)
new_roidb['seg_areas'] = seg_areas
time_end = time.time()
time_used = time_end - time_start
print("Maskrcnn 裁剪耗时{}".format(time_used))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
else:
roidb_copy = creat_new_roidb_empty(roidb[i], img_crop_w,
img_crop_h)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
blob = blob_utils.im_list_to_blob(processed_img)
return blob, imgs_scale, all_roidb
def get_img_blob(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
roidb_base = {}
all_roidb = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
for i in range(num_images):
# print(i)
roidb_copy = {}
# print("minibatch.py 220 @@@@@@@roidb[boxes]{}".format(roidb[i]['boxes']))
img_path = roidb[i][u'image']
# print("minibatch.py 222 image:------------------",img_path)
img = cv2.imread(img_path)
h, w = img.shape[:2]
if roidb[i]['flipped']: # 如果图的flipped为翻转,则翻转
img = img[:, ::-1, :]
# img=np.array(img)
target_size = cfg.TRAIN.SCALES[scale_inds[i]] # target_size是600
im, imscale = blob_utils.prep_im_for_blob( # im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
img, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
z = crop(h, w, roidb[i])
coor, new_boxes_crop, new_classes_crop, img_crop_w, img_crop_h, index_list = z
if not new_boxes_crop == []:
# print("minibatch 237 ",index_list)
gt_class = roidb[i]['gt_classes'][index_list]
# print("minibatch 239 ",gt_class)
gt_overlaps = roidb[i]['gt_overlaps'][index_list]
max_overlaps = roidb[i]['max_overlaps'][
index_list] #np.max(gt_overlaps,axis=1)
max_classes = roidb[i]['max_classes'][
index_list] #np.argmax(gt_overlaps,axis=1)
roidb_copy = roidb[i]
roidb_copy['boxes'] = np.array(new_boxes_crop, dtype=np.float32)
roidb_copy[u'width'] = img_crop_w
roidb_copy[u'height'] = img_crop_h
roidb_copy['gt_classes'] = gt_class
roidb_copy['gt_overlaps'] = gt_overlaps
roidb_copy['is_crowd'] = roidb[i]['is_crowd'][index_list]
roidb_copy['box_to_gt_ind_map'] = roidb[i]['box_to_gt_ind_map'][
index_list]
roidb_copy[u'max_classes'] = max_classes
# print(roidb_copy[-1]['max_overlaps'])
roidb_copy['max_overlaps'] = max_overlaps
# m=roidb_copy[-1]['max_overlaps']
box_target = compute_bbox_regression_targets(roidb_copy)
roidb_copy['bbox_target'] = box_target
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(
img[coor[0]:coor[1], coor[2]:coor[3]], cfg.PIXEL_MEANS,
target_size, cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
else:
roidb_copy = roidb[i]
roidb_copy[u'width'] = img_crop_w
roidb_copy[u'height'] = img_crop_h
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(
img[coor[0]:coor[1], coor[2]:coor[3]], cfg.PIXEL_MEANS,
target_size, cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
# if not processed_img==[]:
blob = blob_utils.im_list_to_blob(processed_img)
# print(" minibatch 266 @@@@@@@imgs_scale{}, roidb[boxes]{}]".format(imgs_scale,all_roidb))
return blob, imgs_scale, all_roidb
def _get_image_blob_s6(roidb, roidb_noclass1):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
error_flag = [0, 0]
for i in range(num_images):
roidb_noclass = roidb_noclass1.copy()
if roidb[i][u'image'].split('/')[-1] == u'test.jpg': #test.jpg
random_bbox = dict()
random_bbox['kernel_size_x'] = int(WIDTH / 5)
random_bbox['kernel_size_y'] = int(HEIGHT / 5)
random_bbox['tl_x'] = 0
random_bbox['tl_y'] = 0
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox['kernel_size_x']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox['kernel_size_y']
im = cv2.imread(roidb[i]['image'])[y0:y1, x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
# cv2.imwrite('/home/icubic/aa.png',im)
error_flag[i] = 0
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
else:
if 1:
if len(roidb[i][u'boxes']) == 0:
#print('no bbox')
random_bbox = dict()
random_flag = random.randint(0, 1)
real_yuanlai_width = roidb[i][u'width'] * 1
real_yuanlai_height = roidb[i][u'height'] * 1
width_ratio = float(real_yuanlai_width) / WIDTH_FIX #1024
height_after_ratio = int(
float(real_yuanlai_height) / width_ratio)
width_after_ratio = WIDTH_FIX #1024
if 1:
if random_flag == 0:
#print(random_flag)
random_bbox['kernel_size_x'] = int(
float(width_after_ratio) /
5.0) #int(WIDTH / 5.0)
random_bbox['kernel_size_y'] = int(
float(height_after_ratio) /
5.0) #int(HEIGHT / 5.0)
random_X = width_after_ratio - random_bbox[
'kernel_size_x']
random_Y = height_after_ratio - random_bbox[
'kernel_size_y']
try:
random_bbox['tl_x'] = random.randint(
0, random_X)
random_bbox['tl_y'] = random.randint(
0, random_Y)
except:
print('aa')
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox[
'kernel_size_x']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox[
'kernel_size_y']
im = cv2.imread(roidb[i][u'image'])
im = cv2.resize(
im,
(width_after_ratio, height_after_ratio))[y0:y1,
x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
else:
#print(random_flag)
random_bbox['kernel_size_x'] = int(
float(width_after_ratio) / 1.0)
random_bbox['kernel_size_y'] = int(
float(height_after_ratio) / 1.0)
random_X = width_after_ratio - random_bbox[
'kernel_size_x']
random_Y = height_after_ratio - random_bbox[
'kernel_size_y']
random_bbox['tl_x'] = random.randint(0, random_X)
random_bbox['tl_y'] = random.randint(0, random_Y)
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox[
'kernel_size_x']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox[
'kernel_size_y']
im = cv2.imread(roidb[i][u'image'])
im = cv2.resize(
im,
(width_after_ratio, height_after_ratio))[y0:y1,
x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
else:
im = cv2.imread(roidb[i][u'image'])
im = cv2.resize(im, (WIDTH, HEIGHT))
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
# cv2.imwrite('/home/icubic/daily_work/circruit_model/tmp_images/aa.png',im)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i][
'flipped']: #for image flip background training
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
continue
#if len(roidb[i][u'boxes']==1):
# print('roidb=1')
#if len(roidb[i][u'boxes']>1):
# print('roidb>1')
real_yuanlai_width = roidb[i][u'width'] * 1
real_yuanlai_height = roidb[i][u'height'] * 1
width_ratio = float(real_yuanlai_width) / WIDTH_FIX
height_after_ratio = int(
float(real_yuanlai_height) / width_ratio)
width_after_ratio = WIDTH_FIX
real_class = [] #roidb[i]['gt_classes'][0]
num_real_class = len(roidb[i]['gt_classes'])
random_bbox = dict()
random_bbox['kernel_size_x'] = int(
float(width_after_ratio) / 5.0) #int(WIDTH_FIX / 5)
random_bbox['kernel_size_y'] = int(
float(height_after_ratio) / 5.0) #int(WIDTH_FIX / 5)
if 1:
w_tongji = 0
h_tongji = 0
for i_tongji, sub_boxes_tongji in enumerate(
roidb[i][u'boxes']):
crop_x0_tongji = int(sub_boxes_tongji[0] /
real_yuanlai_width *
width_after_ratio)
crop_y0_tongji = int(sub_boxes_tongji[1] /
real_yuanlai_height *
height_after_ratio)
crop_x1_tongji = int(sub_boxes_tongji[2] /
real_yuanlai_width *
width_after_ratio)
crop_y1_tongji = int(sub_boxes_tongji[3] /
real_yuanlai_height *
height_after_ratio)
w_tongji = crop_x1_tongji - crop_x0_tongji
h_tongji = crop_y1_tongji - crop_y0_tongji
if w_tongji > int(WIDTH_FIX / 5.0) or h_tongji > int(
WIDTH_FIX / 5.0):
random_bbox['kernel_size_x'] = int(
float(width_after_ratio) / 1.0)
random_bbox['kernel_size_y'] = int(
float(height_after_ratio) / 1.0)
break
if random_bbox['kernel_size_x'] == int(WIDTH_FIX / 5.0):
random_X = width_after_ratio - random_bbox['kernel_size_x']
random_Y = height_after_ratio - random_bbox['kernel_size_y']
random_bbox['tl_x'] = random.randint(0, random_X)
random_bbox['tl_y'] = random.randint(0, random_Y)
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox['kernel_size_x']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox['kernel_size_y']
try:
im = cv2.imread(roidb[i][u'image'])
except:
im = cv2.imread(roidb[i][u'image'])
im = cv2.resize(
im, (width_after_ratio, height_after_ratio))[y0:y1,
x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
sum_inside_overlaps = 0
boxes_inside_overlaps = []
for i_roidb, sub_boxes in enumerate(roidb[i][u'boxes']):
crop_x0 = int(sub_boxes[0] / real_yuanlai_width *
width_after_ratio)
crop_y0 = int(sub_boxes[1] / real_yuanlai_height *
height_after_ratio)
crop_x1 = int(sub_boxes[2] / real_yuanlai_width *
width_after_ratio)
crop_y1 = int(sub_boxes[3] / real_yuanlai_height *
height_after_ratio)
#real_x0 = float(crop_x0 - x0)*1024/224 # float(crop_x0) / 1024 * 224
#real_y0 = float(crop_y0 - y0)*1024/224 # float(crop_y0) / 1024 * 224
#real_x1 = float(crop_x1 - x0)*1024/224 # float(crop_x1) / 1024 * 224
#real_y1 = float(crop_y1 - y0)*1024/224
overlaps_rate = solve_coincide(
(x0, y0, x1, y1),
(crop_x0, crop_y0, crop_x1, crop_y1))
if overlaps_rate > 0.9:
sum_inside_overlaps = sum_inside_overlaps + 1
#real_x0 = crop_x0 - x0 # float(crop_x0) / 1024 * 224
#real_y0 = crop_y0 - y0 # float(crop_y0) / 1024 * 224
#real_x1 = crop_x1 - x0 # float(crop_x1) / 1024 * 224
#real_y1 = crop_y1 - y0
real_x0 = float(crop_x0 - x0) * WIDTH / (
random_bbox['kernel_size_x']
) # float(crop_x0) / 1024 * 224
real_y0 = float(crop_y0 - y0) * HEIGHT / (
random_bbox['kernel_size_y']
) # float(crop_y0) / 1024 * 224
real_x1 = float(crop_x1 - x0) * WIDTH / (
random_bbox['kernel_size_x']
) # float(crop_x1) / 1024 * 224
real_y1 = float(crop_y1 - y0) * HEIGHT / (
random_bbox['kernel_size_y'])
if real_x0 < 0:
real_x0 = 0
if real_x0 > WIDTH:
real_x0 = WIDTH
if real_x1 < 0:
real_x1 = 0
if real_x1 > WIDTH:
real_x1 = WIDTH
if real_y0 < 0:
real_y0 = 0
if real_y0 > HEIGHT:
real_y0 = HEIGHT
if real_y1 < 0:
real_y1 = 0
if real_y1 > HEIGHT:
real_y1 = HEIGHT
#cv2.rectangle(im, (int(real_x0), int(real_y0)), (int(real_x1), int(real_y1)), (0, 255, 255), 3)
#cv2.imwrite('/home/icubic/daily_work/code/Detectron/detectron/datasets/data/shanghai/aa.png',im)
boxes_inside_overlaps.append(
[real_x0, real_y0, real_x1, real_y1])
real_class.append(roidb[i]['gt_classes'][i_roidb])
#cv2.rectangle(im, (int(real_x0), int(real_y0)),
#(int(real_x1), int(real_y1)), (255, 0, 255))
#cv2.imwrite('/home/icubic/daily_work/code/circruit/new/result/uu.png', im)
#a = roidb[i]['gt_overlaps'].toarray()
if sum_inside_overlaps > 0:
num_valid_objs = sum_inside_overlaps * 1
boxes = np.zeros((num_valid_objs, 4), dtype=np.float32)
gt_classes = np.zeros((num_valid_objs), dtype=np.int32)
gt_overlaps = np.zeros((num_valid_objs, REAL_CLASS),
dtype=np.float32)
box_to_gt_ind_map = np.zeros((num_valid_objs),
dtype=np.int32)
is_crowd = np.zeros((num_valid_objs), dtype=np.bool)
for ix in range(num_valid_objs):
gt_classes[ix] = real_class[ix] #real_class*1
try:
gt_overlaps[ix, real_class] = 1.0
except:
print('error')
is_crowd[ix] = False
box_to_gt_ind_map[ix] = ix
for i_index in range(4):
boxes[ix, i_index] = boxes_inside_overlaps[ix][
i_index]
#for ix in range(num_valid_objs):
#box_to_gt_ind_map[ix] = ix
#cls = real_class*1
roidb_noclass['boxes'] = np.append(
roidb_noclass['boxes'], boxes, axis=0)
roidb_noclass['gt_classes'] = np.append(
roidb_noclass['gt_classes'], gt_classes)
#mm = np.append(
# roidb_noclass['gt_overlaps'].toarray(), gt_overlaps,axis=0)
roidb_noclass['gt_overlaps'] = np.append(
roidb_noclass['gt_overlaps'].toarray(),
gt_overlaps)
roidb_noclass['gt_overlaps'] = scipy.sparse.csr_matrix(
roidb_noclass['gt_overlaps'])
#mm = np.append(mm, gt_overlaps, axis=0)
#roidb_noclass['gt_overlaps'] = scipy.sparse.csr_matrix(mm)
roidb_noclass['is_crowd'] = np.append(
roidb_noclass['is_crowd'], is_crowd)
roidb_noclass['box_to_gt_ind_map'] = np.append(
roidb_noclass['box_to_gt_ind_map'],
box_to_gt_ind_map)
gt_overlaps = roidb_noclass['gt_overlaps'].toarray()
# max overlap with gt over classes (columns)
max_overlaps = gt_overlaps.max(axis=1)
# gt class that had the max overlap
max_classes = gt_overlaps.argmax(axis=1)
roidb_noclass['max_classes'] = max_classes
roidb_noclass['max_overlaps'] = max_overlaps
# sanity checks
# if max overlap is 0, the class must be background (class 0)
zero_inds = np.where(max_overlaps == 0)[0]
assert all(max_classes[zero_inds] == 0)
# if max overlap > 0, the class must be a fg class (not class 0)
nonzero_inds = np.where(max_overlaps > 0)[0]
assert all(max_classes[nonzero_inds] != 0)
roidb_noclass[
'bbox_targets'] = compute_bbox_regression_targets(
roidb_noclass)
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
else:
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
#print('aa')
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
else:
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales, error_flag
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
#print("qwryyudfggjhgkjgk{}sdfgfdsg".format(roidb))
# print("minibatch.py 118 此轮roidb的长度为:",len(roidb))
# 获得数据的数量
num_images = len(roidb)
# 在这批数据中最XXXX的采样随机比例 (600,)这个的len()只是1 np.random.randint即在第一个参数和'high'值-1,的整数间生成list,list元素数量是roidb的数量
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES),
size=num_images) # 因为cfg.TRAIN.SCALES为(600,), 所以生成的是一个全为0的list
processed_ims = [] # 图片数据的list
im_scales = [] # 尺度比例list
for i in range(num_images):
try:
im = cv2.imread(roidb[i]['image']) # 读取roidb的图像数据
except Exception as e:
print("Exception:", e)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image']) # 如果图像不存在则assert
if roidb[i]['flipped']: # 如果图的flipped为翻转,则翻转
im = im[:, ::-1, :] # 对图像的宽做倒序 即x轴flipped
img_ = im
#-------------在无多尺度版本中增加色彩模块----------------------------------
if MULTI_COLOR:
color_random = random.uniform(0, 1)
#print("color_random--------------",color_random)
if color_random <= GAUSSIAN:
#print("高斯")
img_ = randomGaussian(img_, GAUSSIAN_KERNEL, GAUSSIAN_SIGMAX)
elif color_random > GAUSSIAN and color_random <= CAB:
#print("融合暗")
img_ = Contrast_and_Brightness(ALPHA, GAMMA, img_)
elif color_random > CAB and color_random <= GAMMA:
#print("GAMMA")
img_ = gamma_trans(img_, GAMMA_THRESHOLD)
elif color_random > GAMMA and color_random <= HSV:
#print("HSV")
img_ = img2darker(img_, HSV_THRESHOLD)
elif color_random > HSV and color_random <= BRIGHT:
#print("BRIGHT")
img_ = bright_trans(img_, BRIGHT_THRESHOLD)
else:
pass
#-----------------------------------------------------------------------
# s_t1 = time.time()
target_size = cfg.TRAIN.SCALES[scale_inds[i]] # target_size是600
im, im_scale = blob_utils.prep_im_for_blob( # im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale) # 加入到im_scales的list中
processed_ims.append(im) # 加入到图像数据list中
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
#e_t1 = time.time()
#t_t1 = e_t1 - s_t1
#print("时间_最终：{}".format(t_t1))
return blob, im_scales
def _get_img_blob_mul(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
all_roidb = []
for i in range(num_images):
# print(i)
#print("minibatch.py 187 @@@@@@@ roidb[boxes]{}".format(roidb[i]['boxes']))
img_path = roidb[i]['image']
#print("minibatch.py 189 image:------",img_path)
img = cv2.imread(img_path)
h, w = img.shape[:2]
# 如果图的flipped为翻转,则翻转
if roidb[i]['flipped']:
img = img[:, ::-1, :]
# target_size是600,或者yaml中设置的尺度
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
# s_t = time.time()
# ---------------------尽量不改变原roidb的值
roidb_n = roidb[i]
#
if ROTATION: # 是否旋转
rot_random = random.uniform(0, 1)
if rot_random <= ROT:
roidb_n = {}
roidb_n['boxes'] = copy.deepcopy(roidb[i]['boxes'])
roidb_n['width'] = roidb[i]['width']
roidb_n['height'] = roidb[i]['height']
roidb_n['has_visible_keypoints'] = roidb[i][
'has_visible_keypoints']
roidb_n['flipped'] = roidb[i]['flipped']
roidb_n['seg_areas'] = copy.deepcopy(roidb[i]['seg_areas'])
roidb_n['dataset'] = roidb[i]['dataset']
roidb_n['segms'] = copy.deepcopy(roidb[i]['segms'])
roidb_n['id'] = roidb[i]['id']
roidb_n['image'] = roidb[i]['image']
roidb_n['gt_classes'] = roidb[i]['gt_classes']
roidb_n['gt_overlaps'] = roidb[i]['gt_overlaps']
roidb_n['is_crowd'] = roidb[i]['is_crowd']
roidb_n['box_to_gt_ind_map'] = roidb[i]['box_to_gt_ind_map']
roidb_n['max_classes'] = roidb[i]['max_classes']
roidb_n['max_overlaps'] = roidb[i]['max_overlaps']
roidb_n['bbox_targets'] = roidb[i]['bbox_targets']
if ROT_RANDOM:
angle = random.randint(-ROT_RANDOM_ANGLE, ROT_RANDOM_ANGLE)
else:
angle = random.randint(1, 4)
angle = angle * 90.0
img_, h_, w_, l, n_w, n_h = rotate_image(img_, angle,
True) # 裁剪完后的图
angle = (angle / 180.0 * math.pi)
# 坐标调整，先计算ｐａｄ后坐标，再计算旋转后，在计算裁剪后
for m, roi_boxes in enumerate(roidb_n['boxes']):
roi_boxes[0] = roi_boxes[0] + int((l - w_) / 2.0)
roi_boxes[1] = roi_boxes[1] + int((l - h_) / 2.0)
roi_boxes[2] = roi_boxes[2] + int((l - w_) / 2.0)
roi_boxes[3] = roi_boxes[3] + int((l - h_) / 2.0)
x0, y0 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[1], angle)
x1, y1 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[1], angle)
x2, y2 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[3], angle)
x3, y3 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[3], angle)
min_x, min_y, max_x, max_y = box_coordinate(
x0, y0, x1, y1, x2, y2, x3, y3)
min_x = min_x - int((l - n_w) / 2.0)
min_y = min_y - int((l - n_h) / 2.0)
max_x = max_x - int((l - n_w) / 2.0)
max_y = max_y - int((l - n_h) / 2.0)
roidb_n['boxes'][m][0] = int(min_x)
roidb_n['boxes'][m][1] = int(min_y)
roidb_n['boxes'][m][2] = int(max_x)
roidb_n['boxes'][m][3] = int(max_y)
# Maskrcnn的点的扩增模块----------------------------------------------------------------
if MASKRCNN_SWITCH:
for m_segme, roi_segms in enumerate(roidb_n['segms']):
for m_segme_index in range(len(roi_segms) / 2):
roi_segms[m_segme][
m_segme_index] = roi_segms_point + int(
(l - w_) / 2.0)
roi_segms[m_segme][m_segme_index +
1] = roi_segms_point + int(
(l - h_) / 2.0)
x_segme, y_segme = rot_compute(
l / 2, l / 2,
roi_segms[m_segme][m_segme_index],
roi_segms[m_segme][m_segme_index + 1], angle)
roidb_n['segms'][m_segme][m_segme_index] = int(
x_segme)
roidb_n['segms'][m_segme][m_segme_index +
1] = int(y_segme)
# -----------------------------------------------------------------------------------
# roidb[i] = roidb_n
else:
pass
# e_t_ = time.time()
# t_t_ = e_t_ - s_t
# print("旋转时间_：{}".format(t_t_))
coor, new_boxes_crop, img_crop_w, img_crop_h, index_list = crop(
h, w, roidb_n, WIDTH_REGION, HEIGHT_REGION)
# 裁剪后实例分割的相对应的segms -------------Maskrcnn的点的扩增模块-----------------------------------
if MASKRCNN_SWITCH:
roidb_n['segms'] = roidb_n['segms'][index_list]
segms = []
seg_areas = []
for m_s, roi_s in enumerate(roidb_n['segms']):
roi_s = np.array(roi_s).reshape(1, -1, 2)
img_se = np.zeros((roidb_n['height'], roidb_n['width'], 3))
imag = cv2.drawContours(img_se, roi_s, -1, (255, 255, 255), -1)
imag = imag[coor[0]:coor[1], coor[2]:coor[3]]
imag = cv2.cvtColor(imag, cv2.COLOR_BGR2GRAY)
_, contours, hierarchy = cv2.findContours(
imag, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
for kk, con_point in enumerate(contours):
if con_point[0][0] == 0:
con_point[0][0] = 1
elif con_point[0][0] == img_crop_w - 1:
con_point[0][0] = con_point[0][0] - 1
else:
pass
if con_point[0][1] == 0:
con_point[0][1] = 1
elif con_point[0][1] == img_crop_w - 1:
con_point[0][1] = con_point[0][0] - 1
else:
pass
roi_s = contours.reshape(1, -1)
eara = float(int(cv2.contourArea(contours[0])))
segms.append(roi_s)
seg_areas.append(eara)
roidb_n['segms'] = segms
roidb_n['seg_areas'] = np.array(eara, dtype=np.float)
# -------------------------------------------------------------------------------------------------------
# 深拷贝以避免改变原始图像数据
img_ = copy.deepcopy(img_[coor[0]:coor[1], coor[2]:coor[3]])
# e_t = time.time()
# t_t = e_t - s_t
# print("时间：{}".format(t_t))
# 随机决定要不要做色度上的数据扩增
if MULTI_COLOR:
color_random = random.uniform(0, 1)
#print("ccccccc",color_random)
if color_random <= GAUSSIAN:
#print("高斯")
img_ = randomGaussian(img_, GAUSSIAN_KERNEL, GAUSSIAN_SIGMAX)
elif color_random <= CAB:
pass
else:
#print("融合暗")
img_ = Contrast_and_Brightness(ALPHA, GAMMA, img_)
if not new_boxes_crop == []:
#----如果裁剪完后有缺陷
#print("执行了minibatch.py 230")
new_roidb = creat_new_roidb(roidb_n, img_crop_w, img_crop_h,
new_boxes_crop, index_list)
# e_t2 = time.time()
# t_t2= e_t2 - s_t
#print("时间_创建新roidb：{}".format(t_t2))
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
#im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
# e_t3 = time.time()
# t_t3= e_t3 - s_t
# print("时间_创建新roidb：{}".format(t_t3))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
else:
#----如果裁剪完后是一张好的图
#print("执行了minibatch.py 239")
roidb_copy = creat_new_roidb_empty(roidb_n, img_crop_w, img_crop_h)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
blob = blob_utils.im_list_to_blob(processed_img)
# print(" minibatch 266 @@@@@@@imgs_scale{}, roidb[boxes]{}]".format(imgs_scale,all_roidb))
#print(len(all_roidb))
#print(all_roidb)
# e_t__ = time.time()
# t_t__ = e_t__ - s_t
# print("时间_最终：{}".format(t_t__))
return blob, imgs_scale, all_roidb
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
#print("qwryyudfggjhgkjgk{}sdfgfdsg".format(roidb))
# print("minibatch.py 118 此轮roidb的长度为:",len(roidb))
# 获得数据的数量
num_images = len(roidb)
# 在这批数据中最XXXX的采样随机比例 (600,)这个的len()只是1 np.random.randint即在第一个参数和'high'值-1,的整数间生成list,list元素数量是roidb的数量
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES),
size=num_images) # 因为cfg.TRAIN.SCALES为(600,), 所以生成的是一个全为0的list
processed_ims = [] # 图片数据的list
im_scales = [] # 尺度比例list
for i in range(num_images):
#print("minibatch.py 109 @@@@@@@ roidb的boxes{}".format(roidb[i]['boxes']))
try:
im = cv2.imread(roidb[i]['image']) # 读取roidb的图像数据
#print("---这是此次的for循环的第{}/{}轮,图像为{}".format(i,num_images,roidb[i]['image']))
except Exception as e:
print("Exception:", e)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image']) # 如果图像不存在则assert
if roidb[i]['flipped']: # 如果图的flipped为翻转,则翻转
im = im[:, ::-1, :] # 对图像的宽做倒序 即x轴flipped
# s_t1 = time.time()
target_size = cfg.TRAIN.SCALES[scale_inds[i]] # target_size是600
im, im_scale = blob_utils.prep_im_for_blob( # im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale) # 加入到im_scales的list中
processed_ims.append(im) # 加入到图像数据list中
#--------------------- 无多尺度下开启旋转模块 --------------------------------------
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
if ROTATION: # 是否旋转
rot_random = random.uniform(0, 1)
if rot_random <= ROT:
roidb_n = {}
roidb_n['boxes'] = copy.deepcopy(roidb[i]['boxes'])
roidb_n['width'] = roidb[i]['width']
roidb_n['height'] = roidb[i]['height']
roidb_n['has_visible_keypoints'] = roidb[i][
'has_visible_keypoints']
roidb_n['flipped'] = roidb[i]['flipped']
roidb_n['seg_areas'] = roidb[i]['seg_areas']
roidb_n['dataset'] = roidb[i]['dataset']
roidb_n['segms'] = roidb[i]['segms']
roidb_n['id'] = roidb[i]['id']
roidb_n['image'] = roidb[i]['image']
roidb_n['gt_classes'] = roidb[i]['gt_classes']
roidb_n['gt_overlaps'] = roidb[i]['gt_overlaps']
roidb_n['is_crowd'] = roidb[i]['is_crowd']
roidb_n['box_to_gt_ind_map'] = roidb[i]['box_to_gt_ind_map']
roidb_n['max_classes'] = roidb[i]['max_classes']
roidb_n['max_overlaps'] = roidb[i]['max_overlaps']
roidb_n['bbox_targets'] = roidb[i]['bbox_targets']
if ROT_RANDOM:
angle = random.randint(-ROT_RANDOM_ANGLE, ROT_RANDOM_ANGLE)
else:
angle = random.randint(1, 4)
angle = angle * 90.0
img_, h_, w_, l, n_w, n_h = rotate_image(img_, angle,
True) # 裁剪完后的图
angle = (angle / 180.0 * math.pi)
# 坐标调整，先计算ｐａｄ后坐标，再计算旋转后，在计算裁剪后
for m, roi_boxes in enumerate(roidb_n['boxes']):
roi_boxes[0] = roi_boxes[0] + int((l - w_) / 2.0)
roi_boxes[1] = roi_boxes[1] + int((l - h_) / 2.0)
roi_boxes[2] = roi_boxes[2] + int((l - w_) / 2.0)
roi_boxes[3] = roi_boxes[3] + int((l - h_) / 2.0)
x0, y0 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[1], angle)
x1, y1 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[1], angle)
x2, y2 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[3], angle)
x3, y3 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[3], angle)
min_x, min_y, max_x, max_y = box_coordinate(
x0, y0, x1, y1, x2, y2, x3, y3)
min_x = min_x - int((l - n_w) / 2.0)
min_y = min_y - int((l - n_h) / 2.0)
max_x = max_x - int((l - n_w) / 2.0)
max_y = max_y - int((l - n_h) / 2.0)
roidb_n['boxes'][m][0] = int(min_x)
roidb_n['boxes'][m][1] = int(min_y)
roidb_n['boxes'][m][2] = int(max_x)
roidb_n['boxes'][m][3] = int(max_y)
# Maskrcnn的点的扩增模块----------------------------------------------------------------
if MASKRCNN_SWITCH:
for m_segme, roi_segms in enumerate(roidb_n['segms']):
for m_segme_index in range(len(roi_segms) / 2):
roi_segms[m_segme][
m_segme_index] = roi_segms_point + int(
(l - w_) / 2.0)
roi_segms[m_segme][m_segme_index +
1] = roi_segms_point + int(
(l - h_) / 2.0)
x_segme, y_segme = rot_compute(
l / 2, l / 2,
roi_segms[m_segme][m_segme_index],
roi_segms[m_segme][m_segme_index + 1], angle)
roidb_n['segms'][m_segme][m_segme_index] = int(
x_segme)
roidb_n['segms'][m_segme][m_segme_index +
1] = int(y_segme)
# -----------------------------------------------------------------------------------
roidb[i] = roidb_n
else:
pass
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
#im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
# e_t3 = time.time()
# t_t3= e_t3 - s_t
# print("时间_创建新roidb：{}".format(t_t3))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
#--------------------- 无多尺度下开启旋转模块 --------------------------------------
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
#e_t1 = time.time()
#t_t1 = e_t1 - s_t1
#print("时间_最终：{}".format(t_t1))
return blob, im_scales
def _get_img_blob_mul(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
all_roidb = []
for i in range(num_images):
img_path = roidb[i]['image']
img = cv2.imread(img_path)
h, w = img.shape[:2]
# 如果图的flipped为翻转,则翻转
if roidb[i]['flipped']:
img = img[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
#print("roidb:---------------------",roidb[i])
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
#----------------------旋转模块-------------------------------
# ---------------------尽量不改变原roidb的值
roidb_n = roidb[i]
if ROTATION: # 是否旋转
rot_random = random.uniform(0, 1)
if rot_random <= ROT:
if ROT_RANDOM:
angle = random.randint(-ROT_RANDOM_ANGLE, ROT_RANDOM_ANGLE)
else:
angle = random.randint(1, 4)
angle = angle * 90.0
img_, h_, w_, l, n_w, n_h = rotate_image(img_, angle,
True) # 裁剪完后的图
angle = (angle / 180.0 * math.pi)
roidb_n = creat_rotation_roidb(roidb_n, angle, l, w_, h_, n_w,
n_h, MASKRCNN_SWITCH)
else:
pass
# e_t_ = time.time()
# t_t_ = e_t_ - s_t
# print("旋转时间_：{}".format(t_t_))
# print("copy图像时间：{}".format(t_t_))
coor, new_boxes_crop, img_crop_w, img_crop_h, index_list = crop(
h, w, roidb[i], WIDTH_REGION, HEIGHT_REGION)
# 深拷贝以避免改变原始图像数据
img_ = copy.deepcopy(img_[coor[0]:coor[1], coor[2]:coor[3]])
# cv2.imwrite("crop_image.jpg", img_)
# ------------------------ 色彩模块 --------------------------
if MULTI_COLOR:
color_random = random.uniform(0, 1)
#print("ccccccc",color_random)
if color_random <= GAUSSIAN:
#print("高斯")
img_ = randomGaussian(img_, GAUSSIAN_KERNEL, GAUSSIAN_SIGMAX)
elif color_random > GAUSSIAN and color_random <= CAB:
#print("融合暗")
img_ = Contrast_and_Brightness(ALPHA, GAMMA, img_)
elif color_random > CAB and color_random <= GAMMA:
#print("GAMMA")
img_ = gamma_trans(img_, GAMMA_THRESHOLD)
elif color_random > GAMMA and color_random <= HSV:
#print("GAMMA")
img_ = img2darker(img_, HSV_THRESHOLD)
elif color_random > HSV and color_random <= BRIGHT:
#print("GAMMA")
img_ = bright_trans(img_, BRIGHT_THRESHOLD)
else:
pass
#----------------------------------------------------------
if not new_boxes_crop == []:
time_start = time.time()
new_roidb = creat_new_roidb(roidb[i], img_crop_w, img_crop_h,
new_boxes_crop, index_list, coor,
MASKRCNN_SWITCH) #MASKRCNN_SWITCH)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
# -----maskrcnn相关的的实例分割的相关代码已集成到creat_new_roidb模块中
# time_start = time.time()
# if MASKRCNN_SWITCH :
# segms = []
# seg_areas = []
# #print("roidb:---------------------",new_roidb )
# if len(new_roidb['segms'])>0:
# for m_s, roi_s in enumerate(new_roidb['segms']):
# roi_s = np.array(roi_s[0][:-4]).reshape(1,-1,2)
# img_se = np.zeros((roidb[i]['height'], roidb[i]['width'], 3))
# #img_se = np.zeros((roidb[i]['height'], roidb[i]['width']))
# img_se_ = img_se.copy()[coor[0]:coor[1], coor[2]:coor[3]]
# imag = cv2.drawContours(img_se, [roi_s],-1,(255,255,255),-1)
# #cv2.imwrite("mask_raw.jpg", imag)
# imag = imag[coor[0]:coor[1], coor[2]:coor[3]]
# # 必须转化为uint8格式,才可以进行二值化
# imag = np.array(imag, dtype = np.uint8)
# #cv2.imwrite("mask_single.jpg", imag)
# #imag = cv2.imread("mask_single.jpg", 3)
# print("!@#^%^&^*^*",imag.dtype)
# gray = cv2.cvtColor(imag, cv2.COLOR_BGR2GRAY)
# # 一定要做二值化,否则出来的是散点
# ret, binary = cv2.threshold(gray,127,255,cv2.THRESH_BINARY)
# _, contours, hierarchy = cv2.findContours(binary,cv2.RETR_TREE,cv2.CHAIN_APPROX_SIMPLE)
# img_fill = cv2.drawContours(imag, contours,-1,(0,255,0),-1)
# #cv2.imwrite("mask_single_fill.jpg", img_fill)
# if len(contours) > 0:
# #print("------------------contours:-----------------",contours)
# for kk, con_point in enumerate(contours):
# if con_point[0][0][0] == 0:
# con_point[0][0][0] = 1
# elif con_point[0][0][0] == img_crop_w -1:
# con_point[0][0][0] = img_crop_w - 2
# else: pass
# if con_point[0][0][1] == 0:
# con_point[0][0][1] = 1
# elif con_point[0][0][1] == img_crop_h -1:
# con_point[0][0][1] = img_crop_h - 2
# else: pass
# roi_s = np.array(contours).reshape(1,-1)
# area = float(int(cv2.contourArea(contours[0])))
# segms.append(roi_s.tolist())
# seg_areas.append(area)
# # segms_ = []
# # # 验证mask框-----
# # for rk, roi_k in enumerate(segms):
# # roi_k = np.array(roi_k).reshape(-1,1,2)
# # segms_.append(roi_k)
# # #print("!!!!!points!!!!!!!",segms_)
# # imag_ = cv2.drawContours(img_se_, segms_ ,-1,(255,255,255),-1)
# # cv2.imwrite("maskrcnn_crop.jpg", imag_)
# # ---------------------------------------------------------------------------------
# new_roidb['segms'] = segms
# new_roidb['seg_areas'] = np.array(seg_areas, dtype = np.float32)
time_end = time.time()
time_used = time_end - time_start
print("Maskrcnn 裁剪耗时{}".format(time_used))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
else:
roidb_copy = creat_new_roidb_empty(roidb[i], img_crop_w,
img_crop_h)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
blob = blob_utils.im_list_to_blob(processed_img)
return blob, imgs_scale, all_roidb
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
#print("qwryyudfggjhgkjgk{}sdfgfdsg".format(roidb))
# print("minibatch.py 118 此轮roidb的长度为:",len(roidb))
# 获得数据的数量
num_images = len(roidb)
# 在这批数据中最XXXX的采样随机比例 (600,)这个的len()只是1 np.random.randint即在第一个参数和'high'值-1,的整数间生成list,list元素数量是roidb的数量
scale_inds = np.random.randint(
0, high=len(cfg.TRAIN.SCALES),
size=num_images) # 因为cfg.TRAIN.SCALES为(600,), 所以生成的是一个全为0的list
processed_img = [] # 图片数据的list
im_scales = []
all_roidb = [] # 尺度比例list
for i in range(num_images):
#print("minibatch.py 109 @@@@@@@ roidb的boxes{}".format(roidb[i]['boxes']))
try:
im = cv2.imread(roidb[i]['image']) # 读取roidb的图像数据
img = im.copy()
#print("---这是此次的for循环的第{}/{}轮,图像为{}".format(i,num_images,roidb[i]['image']))
except Exception as e:
print("Exception:", e)
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image']) # 如果图像不存在则assert
if roidb[i]['flipped']: # 如果图的flipped为翻转,则翻转
im = im[:, ::-1, :] # 对图像的宽做倒序 即x轴flipped
# s_t1 = time.time()
target_size = cfg.TRAIN.SCALES[scale_inds[i]] # target_size是600
im, im_scale = blob_utils.prep_im_for_blob( # im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale) # 加入到im_scales的list中
processed_img.append(im) # 加入到图像数据list中
all_roidb.append(roidb[i])
#--------------------- 无多尺度下开启旋转模块 --------------------------------------
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
roidb_n = roidb[i]
if ROTATION: # 是否旋转
rot_random = random.uniform(0, 1)
if rot_random <= ROT:
if ROT_RANDOM:
angle = random.randint(-ROT_RANDOM_ANGLE, ROT_RANDOM_ANGLE)
else:
angle = random.randint(1, 4)
angle = angle * 90.0
img_, h_, w_, l, n_w, n_h = rotate_image(img_, angle,
True) # 裁剪完后的图
angle = (angle / 180.0 * math.pi)
roidb_n = creat_rotation_roidb(roidb_n, angle, l, w_, h_, n_w,
n_h, MASKRCNN_SWITCH)
else:
pass
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
#im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
# e_t3 = time.time()
# t_t3= e_t3 - s_t
# print("时间_创建新roidb：{}".format(t_t3))
processed_img.append(im)
im_scales.append(imscale)
all_roidb.append(roidb_n)
#--------------------- 无多尺度下开启旋转模块 --------------------------------------
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_img)
#e_t1 = time.time()
#t_t1 = e_t1 - s_t1
#print("时间_最终：{}".format(t_t1))
return blob, im_scales
def _get_image_blob(roidb):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
im_crops = []
for i in range(num_images):
im = cv2.imread(roidb[i]['image'])
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
if cfg.WSL.USE_DISTORTION:
hsv = cv2.cvtColor(im, cv2.COLOR_BGR2HSV)
s0 = npr.random() * (cfg.WSL.SATURATION - 1) + 1
s1 = npr.random() * (cfg.WSL.EXPOSURE - 1) + 1
s0 = s0 if npr.random() > 0.5 else 1.0 / s0
s1 = s1 if npr.random() > 0.5 else 1.0 / s1
hsv = np.array(hsv, dtype=np.float32)
hsv[:, :, 1] = np.minimum(s0 * hsv[:, :, 1], 255)
hsv[:, :, 2] = np.minimum(s1 * hsv[:, :, 2], 255)
hsv = np.array(hsv, dtype=np.uint8)
im = cv2.cvtColor(hsv, cv2.COLOR_HSV2BGR)
if cfg.WSL.USE_CROP:
im_shape = np.array(im.shape)
crop_dims = im_shape[:2] * cfg.WSL.CROP
r0 = npr.random()
r1 = npr.random()
s = im_shape[:2] - crop_dims
s[0] *= r0
s[1] *= r1
im_crop = np.array(
[s[0], s[1], s[0] + crop_dims[0] - 1, s[1] + crop_dims[1] - 1],
dtype=np.int32)
im = im[im_crop[0]:im_crop[2] + 1, im_crop[1]:im_crop[3] + 1, :]
else:
im_crop = np.array([0, 0, im.shape[0] - 1, im.shape[1] - 1],
dtype=np.int32)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
im_crops.append(im_crop)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales, im_crops
def _get_img_blob_mul(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
all_roidb = []
for i in range(num_images):
# print(i)
#print("minibatch.py 187 @@@@@@@ roidb[boxes]{}".format(roidb[i]['boxes']))
img_path = roidb[i]['image']
#print("minibatch.py 189 image:------",img_path)
img = cv2.imread(img_path)
h, w = img.shape[:2]
# 如果图的flipped为翻转,则翻转
if roidb[i]['flipped']:
img = img[:, ::-1, :]
# target_size是600,或者yaml中设置的尺度
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
# s_t = time.time()
if ROTATION: # 是否旋转
rot_random = random.uniform(0, 1)
if rot_random <= ROT:
roidb_n = {}
roidb_n['boxes'] = copy.deepcopy(roidb[i]['boxes'])
roidb_n['width'] = roidb[i]['width']
roidb_n['height'] = roidb[i]['height']
roidb_n['has_visible_keypoints'] = roidb[i][
'has_visible_keypoints']
roidb_n['flipped'] = roidb[i]['flipped']
roidb_n['seg_areas'] = roidb[i]['seg_areas']
roidb_n['dataset'] = roidb[i]['dataset']
roidb_n['segms'] = roidb[i]['segms']
roidb_n['id'] = roidb[i]['id']
roidb_n['image'] = roidb[i]['image']
roidb_n['gt_classes'] = roidb[i]['gt_classes']
roidb_n['gt_overlaps'] = roidb[i]['gt_overlaps']
roidb_n['is_crowd'] = roidb[i]['is_crowd']
roidb_n['box_to_gt_ind_map'] = roidb[i]['box_to_gt_ind_map']
roidb_n['max_classes'] = roidb[i]['max_classes']
roidb_n['max_overlaps'] = roidb[i]['max_overlaps']
roidb_n['bbox_targets'] = roidb[i]['bbox_targets']
if ROT_RANDOM:
angle = random.randint(-ROT_RANDOM_ANGLE, ROT_RANDOM_ANGLE)
else:
angle = random.randint(1, 4)
angle = angle * 90.0
img_, h_, w_, l, n_w, n_h = rotate_image(img_, angle,
True) # 裁剪完后的图
angle = (angle / 180.0 * math.pi)
# 坐标调整，先计算ｐａｄ后坐标，再计算旋转后，在计算裁剪后
for m, roi_boxes in enumerate(roidb_n['boxes']):
roi_boxes[0] = roi_boxes[0] + int((l - w_) / 2.0)
roi_boxes[1] = roi_boxes[1] + int((l - h_) / 2.0)
roi_boxes[2] = roi_boxes[2] + int((l - w_) / 2.0)
roi_boxes[3] = roi_boxes[3] + int((l - h_) / 2.0)
x0, y0 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[1], angle)
x1, y1 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[1], angle)
x2, y2 = rot_compute(l / 2, l / 2, roi_boxes[0],
roi_boxes[3], angle)
x3, y3 = rot_compute(l / 2, l / 2, roi_boxes[2],
roi_boxes[3], angle)
min_x, min_y, max_x, max_y = box_coordinate(
x0, y0, x1, y1, x2, y2, x3, y3)
min_x = min_x - int((l - n_w) / 2.0)
min_y = min_y - int((l - n_h) / 2.0)
max_x = max_x - int((l - n_w) / 2.0)
max_y = max_y - int((l - n_h) / 2.0)
roidb_n['boxes'][m][0] = int(min_x)
roidb_n['boxes'][m][1] = int(min_y)
roidb_n['boxes'][m][2] = int(max_x)
roidb_n['boxes'][m][3] = int(max_y)
roidb[i] = roidb_n
else:
pass
# e_t_ = time.time()
# t_t_ = e_t_ - s_t
# print("旋转时间_：{}".format(t_t_))
coor, new_boxes_crop, img_crop_w, img_crop_h, index_list = crop(
h, w, roidb[i], [0.8, 0.9], [0.8, 0.9])
# 深拷贝以避免改变原始图像数据
img_ = copy.deepcopy(img_[coor[0]:coor[1], coor[2]:coor[3]])
# e_t = time.time()
# t_t = e_t - s_t
# print("时间：{}".format(t_t))
# 随机决定要不要做色度上的数据扩增
if MULTI_COLOR:
color_random = random.uniform(0, 1)
#print("ccccccc",color_random)
if color_random <= GAUSSIAN:
#print("高斯")
img_ = randomGaussian(img_, GAUSSIAN_KERNEL, GAUSSIAN_SIGMAX)
elif color_random <= CAB:
pass
else:
#print("融合暗")
img_ = Contrast_and_Brightness(ALPHA, GAMMA, img_)
if not new_boxes_crop == []:
#print("执行了minibatch.py 230")
new_roidb = creat_new_roidb(roidb[i], img_crop_w, img_crop_h,
new_boxes_crop, index_list)
# e_t2 = time.time()
# t_t2= e_t2 - s_t
#print("时间_创建新roidb：{}".format(t_t2))
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
#im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
# e_t3 = time.time()
# t_t3= e_t3 - s_t
# print("时间_创建新roidb：{}".format(t_t3))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
# img__ = img_
# img_name = str(new_roidb['image']).split("/",-1)[-1]
# for mi, roi_ng in enumerate(new_roidb['boxes']):
# cv2.rectangle(img__, (roi_ng[0], roi_ng[1]),(roi_ng[2], roi_ng[3]), (255,0,0),2)
else:
#print("执行了minibatch.py 239")
roidb_copy = creat_new_roidb_empty(roidb[i], img_crop_w,
img_crop_h)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im, imscale = blob_utils.prep_im_for_blob(img[coor[0]:coor[1],coor[2]:coor[3]], cfg.PIXEL_MEANS, target_size, cfg.TRAIN.MAX_SIZE)
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
# img__ = img_
# img_name = str(roidb_copy['image']).split("/",-1)[-1]
# time_n = time.time()
# cv2.imwrite("/home/icubic/detectron_img/" + img_name[:-4] + "_" +str(time_n) + ".jpg" , img__)
blob = blob_utils.im_list_to_blob(processed_img)
# print(" minibatch 266 @@@@@@@imgs_scale{}, roidb[boxes]{}]".format(imgs_scale,all_roidb))
#print(len(all_roidb))
#print(all_roidb)
# e_t__ = time.time()
# t_t__ = e_t__ - s_t
# print("时间_最终：{}".format(t_t__))
return blob, imgs_scale, all_roidb
def get_img_blob(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
roidb_base = {}
all_roidb = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
for i in range(num_images):
# print(i)
roidb_copy = {}
# print("minibatch.py 220 @@@@@@@roidb[boxes]{}".format(roidb[i]['boxes']))
img_path = roidb[i][u'image']
# print("minibatch.py 222 image:------------------",img_path)
img = cv2.imread(img_path)
# img=np.array(img)
z = crop(img, roidb[i])
image, new_boxes_crop, new_classes_crop, img_crop_w, img_crop_h, index_list = z
# for i in range(len(new_boxes_crop)):
# bbox = new_boxes_crop[i]
# x_min = bbox[0]
# y_min = bbox[1]
# x_max = bbox[2]
# y_max = bbox[3]
# cv2.rectangle(image, (int(x_min), int(y_min)), (int(x_max), int(y_max)), (0, 255, 0), 6)
# cv2.imwrite('./new_img.jpg',image)
if not new_boxes_crop == []:
# print("minibatch 237 ",index_list)
gt_class = roidb[i]['gt_classes'][index_list]
# print("minibatch 239 ",gt_class)
gt_overlaps = roidb[i]['gt_overlaps'][index_list]
max_overlaps = roidb[i]['max_overlaps'][
index_list] #np.max(gt_overlaps,axis=1)
max_classes = roidb[i]['max_classes'][
index_list] #np.argmax(gt_overlaps,axis=1)
roidb_copy = roidb[i]
roidb_copy['boxes'] = np.array(new_boxes_crop, dtype=np.float32)
roidb_copy[u'width'] = img_crop_w
roidb_copy[u'height'] = img_crop_h
roidb_copy['gt_classes'] = gt_class
roidb_copy['gt_overlaps'] = gt_overlaps
roidb_copy['is_crowd'] = roidb[i]['is_crowd'][index_list]
roidb_copy['box_to_gt_ind_map'] = roidb[i]['box_to_gt_ind_map'][
index_list]
roidb_copy[u'max_classes'] = max_classes
# print(roidb_copy[-1]['max_overlaps'])
roidb_copy['max_overlaps'] = max_overlaps
# m=roidb_copy[-1]['max_overlaps']
box_target = compute_bbox_regression_targets(roidb_copy)
roidb_copy['bbox_target'] = box_target
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
else:
roidb_copy = roidb[i]
roidb_copy[u'width'] = img_crop_w
roidb_copy[u'height'] = img_crop_h
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
if not processed_img == []:
blob = blob_utils.im_list_to_blob(processed_img)
# print(" minibatch 266 @@@@@@@imgs_scale{}, roidb[boxes]{}]".format(imgs_scale,all_roidb))
return blob, imgs_scale, all_roidb
def _get_img_blob_mul(roidb):
num_images = len(roidb)
processed_img = []
imgs_scale = []
scale_inds = np.random.randint(0, len(cfg.TRAIN.SCALES), size=num_images)
all_roidb = []
for i in range(num_images):
img_path = roidb[i]['image']
img = cv2.imread(img_path)
h, w = img.shape[:2]
# 如果图的flipped为翻转,则翻转
if roidb[i]['flipped']:
img = img[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
# im是 resize后的数据, im_scale 是resize的比例 参数中cfg.PIXEL_MEANS/cfg.TRAIN.MAX_SIZE在配置文件中有配置
im, imscale = blob_utils.prep_im_for_blob(img, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb[i])
img_ = img.copy() # 其实在原图做完计算之后可以不用考虑值会改变的事了
# print("copy图像时间：{}".format(t_t_))
coor, new_boxes_crop, img_crop_w, img_crop_h, index_list = crop(
h, w, roidb[i], WIDTH_REGION, HEIGHT_REGION)
# 深拷贝以避免改变原始图像数据
img_ = copy.deepcopy(img_[coor[0]:coor[1], coor[2]:coor[3]])
if not new_boxes_crop == []:
new_roidb = creat_new_roidb(roidb[i], img_crop_w, img_crop_h,
new_boxes_crop, index_list)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
time_start = time.time()
if MASKRCNN_SWITCH:
segms = []
seg_areas = []
for m_s, roi_s in enumerate(new_roidb['segms']):
roi_s = np.array(roi_s).reshape(1, -1, 2)
img_se = np.zeros(
(new_roidb['height'], new_roidb['width'], 3))
imag = cv2.drawContours(img_se, roi_s, -1, (255, 255, 255),
-1)
imag = imag[coor[0]:coor[1], coor[2]:coor[3]]
imag = cv2.cvtColor(imag, cv2.COLOR_BGR2GRAY)
_, contours, hierarchy = cv2.findContours(
imag, cv2.RETR_TREE, cv2.CHAIN_APPROX_SIMPLE)
for kk, con_point in enumerate(contours):
if con_point[0][0] == 0:
con_point[0][0] = 1
elif con_point[0][0] == img_crop_w - 1:
con_point[0][0] = con_point[0][0] - 1
else:
pass
if con_point[0][1] == 0:
con_point[0][1] = 1
elif con_point[0][1] == img_crop_w - 1:
con_point[0][1] = con_point[0][0] - 1
else:
pass
roi_s = contours.reshape(1, -1)
eara = float(int(cv2.contourArea(contours[0])))
segms.append(roi_s)
seg_areas.append(eara)
new_roidb['segms'] = segms
new_roidb['seg_areas'] = np.array(eara, dtype=np.float)
time_end = time.time()
time_used = time_end - time_start
print("Maskrcnn 裁剪耗时{}".format(time_used))
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(new_roidb)
else:
roidb_copy = creat_new_roidb_empty(roidb[i], img_crop_w,
img_crop_h)
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, imscale = blob_utils.prep_im_for_blob(img_, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
processed_img.append(im)
imgs_scale.append(imscale)
all_roidb.append(roidb_copy)
blob = blob_utils.im_list_to_blob(processed_img)
return blob, imgs_scale, all_roidb
def _get_image_blob_s6_0(roidb, roidb_noclass1):
"""Builds an input blob from the images in the roidb at the specified
scales.
"""
num_images = len(roidb)
# Sample random scales to use for each image in this batch
scale_inds = np.random.randint(0,
high=len(cfg.TRAIN.SCALES),
size=num_images)
processed_ims = []
im_scales = []
error_flag = [0, 0]
for i in range(num_images):
roidb_noclass = roidb_noclass1.copy()
if roidb[i][u'image'].split('/')[-1] == u'test.jpg':
random_bbox = dict()
random_bbox['kernel_size'] = 224
random_bbox['tl_x'] = 0
random_bbox['tl_y'] = 0
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox['kernel_size']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox['kernel_size']
im = cv2.imread(roidb[i]['image'])[y0:y1, x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
#cv2.imwrite('/home/icubic/aa.png',im)
error_flag[i] = 0
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
else:
if 1:
real_class = [] #roidb[i]['gt_classes'][0]
num_real_class = len(roidb[i]['gt_classes'])
random_bbox = dict()
random_bbox['kernel_size'] = 224
random_bbox['tl_x'] = random.randint(0, 800)
random_bbox['tl_y'] = random.randint(0, 800)
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox['kernel_size']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox['kernel_size']
im = cv2.imread(roidb[i]['image'])[y0:y1, x0:x1]
im = cv2.resize(im, (WIDTH, HEIGHT))
sum_inside_overlaps = 0
boxes_inside_overlaps = []
for i_roidb, sub_boxes in enumerate(roidb[i][u'boxes']):
crop_x0 = int(sub_boxes[0])
crop_y0 = int(sub_boxes[1])
crop_x1 = int(sub_boxes[2])
crop_y1 = int(sub_boxes[3])
#real_x0 = float(crop_x0 - x0)*1024/224 # float(crop_x0) / 1024 * 224
#real_y0 = float(crop_y0 - y0)*1024/224 # float(crop_y0) / 1024 * 224
#real_x1 = float(crop_x1 - x0)*1024/224 # float(crop_x1) / 1024 * 224
#real_y1 = float(crop_y1 - y0)*1024/224
overlaps_rate = solve_coincide(
(x0, y0, x1, y1), (crop_x0, crop_y0, crop_x1, crop_y1))
if overlaps_rate > 0.9:
sum_inside_overlaps = sum_inside_overlaps + 1
#real_x0 = crop_x0 - x0 # float(crop_x0) / 1024 * 224
#real_y0 = crop_y0 - y0 # float(crop_y0) / 1024 * 224
#real_x1 = crop_x1 - x0 # float(crop_x1) / 1024 * 224
#real_y1 = crop_y1 - y0
real_x0 = float(
crop_x0 -
x0) * WIDTH / 224 # float(crop_x0) / 1024 * 224
real_y0 = float(
crop_y0 -
y0) * HEIGHT / 224 # float(crop_y0) / 1024 * 224
real_x1 = float(
crop_x1 -
x0) * WIDTH / 224 # float(crop_x1) / 1024 * 224
real_y1 = float(crop_y1 - y0) * HEIGHT / 224
if real_x0 < 0:
real_x0 = 0
if real_x0 > WIDTH:
real_x0 = WIDTH
if real_x1 < 0:
real_x1 = 0
if real_x1 > WIDTH:
real_x1 = WIDTH
if real_y0 < 0:
real_y0 = 0
if real_y0 > HEIGHT:
real_y0 = HEIGHT
if real_y1 < 0:
real_y1 = 0
if real_y1 > HEIGHT:
real_y1 = HEIGHT
boxes_inside_overlaps.append(
[real_x0, real_y0, real_x1, real_y1])
real_class.append(roidb[i]['gt_classes'][i_roidb])
#cv2.rectangle(im, (int(real_x0), int(real_y0)),
#(int(real_x1), int(real_y1)), (255, 0, 255))
#cv2.imwrite('/home/icubic/daily_work/code/circruit/new/result/uu.png', im)
#a = roidb[i]['gt_overlaps'].toarray()
if sum_inside_overlaps > 0:
num_valid_objs = sum_inside_overlaps * 1
boxes = np.zeros((num_valid_objs, 4), dtype=np.float32)
gt_classes = np.zeros((num_valid_objs), dtype=np.int32)
gt_overlaps = np.zeros((num_valid_objs, 3),
dtype=np.float32)
box_to_gt_ind_map = np.zeros((num_valid_objs),
dtype=np.int32)
is_crowd = np.zeros((num_valid_objs), dtype=np.bool)
for ix in range(num_valid_objs):
gt_classes[ix] = real_class[ix] #real_class*1
try:
gt_overlaps[ix, real_class] = 1.0
except:
print('error')
is_crowd[ix] = False
box_to_gt_ind_map[ix] = ix
for i_index in range(4):
boxes[ix,
i_index] = boxes_inside_overlaps[ix][i_index]
#for ix in range(num_valid_objs):
#box_to_gt_ind_map[ix] = ix
#cls = real_class*1
roidb_noclass['boxes'] = np.append(roidb_noclass['boxes'],
boxes,
axis=0)
roidb_noclass['gt_classes'] = np.append(
roidb_noclass['gt_classes'], gt_classes)
#mm = np.append(
# roidb_noclass['gt_overlaps'].toarray(), gt_overlaps,axis=0)
roidb_noclass['gt_overlaps'] = np.append(
roidb_noclass['gt_overlaps'].toarray(), gt_overlaps)
roidb_noclass['gt_overlaps'] = scipy.sparse.csr_matrix(
roidb_noclass['gt_overlaps'])
#mm = np.append(mm, gt_overlaps, axis=0)
#roidb_noclass['gt_overlaps'] = scipy.sparse.csr_matrix(mm)
roidb_noclass['is_crowd'] = np.append(
roidb_noclass['is_crowd'], is_crowd)
roidb_noclass['box_to_gt_ind_map'] = np.append(
roidb_noclass['box_to_gt_ind_map'], box_to_gt_ind_map)
gt_overlaps = roidb_noclass['gt_overlaps'].toarray()
# max overlap with gt over classes (columns)
max_overlaps = gt_overlaps.max(axis=1)
# gt class that had the max overlap
max_classes = gt_overlaps.argmax(axis=1)
roidb_noclass['max_classes'] = max_classes
roidb_noclass['max_overlaps'] = max_overlaps
# sanity checks
# if max overlap is 0, the class must be background (class 0)
zero_inds = np.where(max_overlaps == 0)[0]
assert all(max_classes[zero_inds] == 0)
# if max overlap > 0, the class must be a fg class (not class 0)
nonzero_inds = np.where(max_overlaps > 0)[0]
assert all(max_classes[nonzero_inds] != 0)
roidb_noclass[
'bbox_targets'] = compute_bbox_regression_targets(
roidb_noclass)
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
else:
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = HEIGHT
roidb[i][u'width'] = WIDTH
if 0:
if sum_inside_overlaps == 0:
roidb[i] = roidb_noclass['0'].copy()
roidb[i][u'height'] = 1024
roidb[i][u'width'] = 1024
if sum_inside_overlaps == 1:
num_valid_objs = 1
roidb[i] = roidb_noclass['1'].copy()
a = roidb[i]['gt_overlaps'].toarray()
#for i_inside in enumerate(sum_inside_overlaps)
if sum_inside_overlaps == 2:
num_valid_objs = 2
roidb[i] = roidb_noclass['2'].copy()
a = roidb[i]['gt_overlaps'].toarray()
if sum_inside_overlaps == 3:
num_valid_objs = 3
roidb[i] = roidb_noclass['3'].copy()
a = roidb[i]['gt_overlaps'].toarray()
if 0:
crop_x0 = int(roidb[i][u'boxes'][0][0])
crop_y0 = int(roidb[i][u'boxes'][0][1])
crop_x1 = int(roidb[i][u'boxes'][0][2])
crop_y1 = int(roidb[i][u'boxes'][0][3])
crop_w = crop_x1 - crop_x0
crop_h = crop_y1 - crop_y0
random_bbox = dict()
random_bbox['kernel_size'] = 224
random_bbox['tl_x'] = random.randint(0, 800)
random_bbox['tl_y'] = random.randint(0, 800)
x0 = random_bbox['tl_x']
x1 = random_bbox['tl_x'] + random_bbox['kernel_size']
y0 = random_bbox['tl_y']
y1 = random_bbox['tl_y'] + random_bbox['kernel_size']
#real_x0 = crop_x0-x0#float(crop_x0) / 1024 * 224
#real_y0 = crop_y0-y0#float(crop_y0) / 1024 * 224
#real_x1 = 1024#float(crop_x1) / 1024 * 224
#real_y1 = 1024#float(crop_y1) / 1024 * 224
overlaps_rate = solve_coincide(
(x0, y0, x1, y1), (crop_x0, crop_y0, crop_x1, crop_y1))
im = cv2.imread(roidb[i]['image'])[y0:y1, x0:x1]
#im = cv2.resize(im, (1024, 1024))
if overlaps_rate > 0.9:
real_x0 = crop_x0 - x0 # float(crop_x0) / 1024 * 224
real_y0 = crop_y0 - y0 # float(crop_y0) / 1024 * 224
real_x1 = crop_x1 - x0 # float(crop_x1) / 1024 * 224
real_y1 = crop_y1 - y0
roidb[i][u'boxes'][0][0] = real_x0
roidb[i][u'boxes'][0][1] = real_y0
roidb[i][u'boxes'][0][2] = real_x1
roidb[i][u'boxes'][0][3] = real_y1
roidb[i][u'height'] = 224
roidb[i][u'width'] = 224
error_flag[i] = 1
#cv2.imwrite('/home/icubic/daily_work/code/Detectron/detectron/datasets/data/s6_test/aa.png',im)
else:
roidb[i] = roidb_noclass.copy()
roidb[i][u'height'] = 224
roidb[i][u'width'] = 224
error_flag[i] = 0
#print('aa')
assert im is not None, \
'Failed to read image \'{}\''.format(roidb[i]['image'])
if roidb[i]['flipped']:
im = im[:, ::-1, :]
target_size = cfg.TRAIN.SCALES[scale_inds[i]]
im, im_scale = blob_utils.prep_im_for_blob(im, cfg.PIXEL_MEANS,
target_size,
cfg.TRAIN.MAX_SIZE)
im_scales.append(im_scale)
processed_ims.append(im)
# Create a blob to hold the input images
blob = blob_utils.im_list_to_blob(processed_ims)
return blob, im_scales, error_flag