def __getitem__(self, idx):
idx = idx % len(self.data)
iD, image, prob = self.data[idx]
#image = (image[:,:,:3 ]*255).astype( np.uint8 )
#image = (image[:,:, 0 ]*255).astype( np.uint8 )
image = np.stack(
(image[:, :, 0], image[:, :, 1] / 2 + image[:, :, 3] / 2,
image[:, :, 2] / 2 + image[:, :, 3] / 2),
axis=-1)
#image = np.stack( ( image[:,:,0], image[:,:,1], image[:,:,2] ), axis=-1 )
#image = utility.to_channels(image, 3)
image = (image * 255).astype(np.uint8)
#print(image.shape, flush=True )
#print(image.min(), image.max(), flush=True )
#assert(False)
obj = ObjectImageTransform(image)
if self.transform:
obj = self.transform(obj)
image = obj.to_value()
#print(image.shape, flush=True )
#assert(False)
return iD, image, prob
def __getitem__(self, idx):
# read image
image, label = self.data[(idx) % len(self.data)]
#A,A_inv = F.compute_norm_mat( image.shape[1], image.shape[0] )
#image = F.equalization(image,A,A_inv)
image = utility.to_channels(image, self.num_channels)
# read background
if self.bbackimage:
idxk = random.randint(1, len(self.databack) - 1)
back = self.databack[idxk]
back = F.resize_image(back,
640,
1024,
resize_mode='crop',
interpolate_mode=cv2.INTER_LINEAR)
back = utility.to_channels(back, self.num_channels)
else:
back = np.ones((640, 1024, 3), dtype=np.uint8) * 255
if self.generate == 'image':
obj = ObjectImageTransform(image)
elif self.generate == 'image_and_mask':
image_org, image_ilu, mask, h = self.ren.generate(image, back)
image_org = utility.to_gray(image_org.astype(np.uint8))
image_org = utility.to_channels(image_org, self.num_channels)
image_org = image_org.astype(np.uint8)
image_ilu = utility.to_gray(image_ilu.astype(np.uint8))
image_ilu = utility.to_channels(image_ilu, self.num_channels)
image_ilu = image_ilu.astype(np.uint8)
mask = mask[:, :, 0]
mask_t = np.zeros((mask.shape[0], mask.shape[1], 2))
mask_t[:, :, 0] = (mask == 0).astype(np.uint8) # 0-backgraund
mask_t[:, :, 1] = (mask == 1).astype(np.uint8)
obj_image = ObjectImageTransform(image_org.copy())
obj_data = ObjectImageAndMaskMetadataTransform(
image_ilu.copy(), mask_t, np.concatenate(
([label], h), axis=0)) #np.array([label])
else:
assert (False)
if self.transform_image:
obj_image = self.transform_image(obj_image)
if self.transform_data:
obj_data = self.transform_data(obj_data)
x_img, y_mask, y_lab = obj_data.to_value()
x_org = obj_image.to_value()
return x_org, x_img, y_mask, y_lab
def transform_image(image, image_size, num_channels):
image = np.array(image, dtype=np.uint8)
image = utility.to_channels(image, num_channels)
obj = ObjectImageTransform(image)
obj = get_transforms_det(image_size)(obj)
# tensor is channel x height x width
# image = image.transpose((2, 0, 1))
# obj.to_tensor()
# value = torch.from_numpy(obj.image).float()
obj.image = obj.image.unsqueeze(0)
return obj.to_dict()
def __getitem__(self, idx):
image = self.data[idx]
Id = idx
image = np.array(image)
image = utility.to_channels(image, self.num_channels)
obj = ObjectImageTransform( image )
if self.transform:
obj = self.transform(obj)
Id = torch.from_numpy( np.array([Id])).float()
return Id, obj.to_value()
def __getitem__(self, idx):
# read image
image, label = self.data[(idx) % len(self.data)]
image = utility.to_channels(image, self.num_channels)
# read background
if self.bbackimage:
idxk = random.randint(1, len(self.databack) - 1)
back = self.databack[idxk] #(idx)%len(self.databack)
back = F.resize_image(back,
640,
1024,
resize_mode='crop',
interpolate_mode=cv2.INTER_LINEAR)
back = utility.to_channels(back, self.num_channels)
else:
back = np.ones((640, 1024, 3), dtype=np.uint8) * 255
if self.generate == 'image':
obj = ObjectImageTransform(image)
elif self.generate == 'image_and_label':
_, image, _ = self.ren.generate(image, back)
image = utility.to_gray(image.astype(np.uint8))
image_t = utility.to_channels(image, self.num_channels)
image_t = image_t.astype(np.uint8)
label = utility.to_one_hot(int(label), self.data.numclass)
obj = ObjectImageAndLabelTransform(image_t, label)
elif self.generate == 'image_and_mask':
_, image, mask = self.ren.generate(image, back)
image = utility.to_gray(image.astype(np.uint8))
image_t = utility.to_channels(image, self.num_channels)
image_t = image_t.astype(np.uint8)
#print( image_t.shape, image_t.min(), image_t.max(), flush=True )
#assert(False)
mask = mask[:, :, 0]
mask_t = np.zeros((mask.shape[0], mask.shape[1], 2))
mask_t[:, :, 0] = (mask == 0).astype(np.uint8) # backgraund
mask_t[:, :, 1] = (mask == 1).astype(np.uint8)
obj = ObjectImageAndMaskMetadataTransform(image_t, mask_t,
np.array([label]))
else:
assert (False)
if self.transform:
obj = self.transform(obj)
return obj.to_dict()
def tranform_image_performs(image,
transform,
num_transform=4,
bsave=False,
bshow=True,
bgrid=False):
frames = []
for i in range(num_transform):
obj = ObjectImageTransform(image)
if bgrid: obj._draw_grid(50, (255, 255, 255))
start = time.time()
obj_transform = transform(obj)
t = time.time() - start
print('{} ({}sec)'.format(transform, t))
image_o = obj.image
image_t = obj_transform.image
frame = np.concatenate((stand(image), stand(image_t)), axis=1)
font = cv2.FONT_HERSHEY_SIMPLEX
txt = 'Transform: + [{} ({:0.4f}sec)]'.format(transform, t)
cv2.putText(frame, txt, (10, 15), font, 0.35, (255, 255, 255), 1,
cv2.LINE_AA)
frames.append(frame)
if bshow:
plt.figure(figsize=(8, 4))
plt.imshow(frame)
plt.title('Transform: +[{} ({:0.4f}sec)]'.format(transform, t))
plt.show()
if bsave:
filename = '../rec/{}'.format(transform)
create_gif(filename, frames, duration=0.5)
print('save: ', filename)
def test_image_transform():
pytvision.transforms.aumentation.interpolate_image_mode = cv2.INTER_LINEAR
image = cv2.imread('../rec/image.jpg')
image = image[:, :, (2, 1, 0)] #[:,:,0]
obj = ObjectImageTransform(image)
trans = transform_aug()
start = time.time()
obj_t = trans(obj)
t = time.time() - start
print('Time {}(s)'.format(t))
image_t = obj_t.image.permute((1, 2, 0)).numpy() #.squeeze(-1)
plt.figure(figsize=(8, 8))
plt.imshow(image_t)
plt.show()