def rgb_processing(self, rgb_img, center, scale, rot, flip, pn, is_train):
rgb_img = crop(rgb_img.copy(), center, scale, [constants.IMG_RES, constants.IMG_RES], rot=rot)
if is_train:
if flip:
rgb_img = flip_img(rgb_img)
rgb_img = np.transpose(rgb_img.astype('float32'),(2,0,1))/255.0
return rgb_img
def rgb_processing(self, rgb_img, center, scale, rot, flip, pn, img_size):
"""Process rgb image and do augmentation."""
if self.is_rotate:
rgb_img = crop_v2(rgb_img, center, scale,
[constants.IMG_RES, constants.IMG_RES], rot=rot)
# flip the image
if flip:
rgb_img = flip_img(rgb_img)
# in the rgb image we add pixel noise in a channel-wise manner
rgb_img[:, :, 0] = np.minimum(255.0, np.maximum(0.0, rgb_img[:, :, 0] * pn[0]))
rgb_img[:, :, 1] = np.minimum(255.0, np.maximum(0.0, rgb_img[:, :, 1] * pn[1]))
rgb_img[:, :, 2] = np.minimum(255.0, np.maximum(0.0, rgb_img[:, :, 2] * pn[2]))
if img_size == 224:
rgb_img_up = rgb_img.copy()
# add color jitter
if self.is_train:
rgb_img_up = color_jitter(rgb_img_up, brightness=0.4, contrast=0.4, saturation=0.4, prob=0.5)
rgb_img_up = rgb_img_up.clip(0, 255)
else:
shape = rgb_img.shape
rgb_img_lr = scipy.misc.imresize(rgb_img, (img_size, img_size), interp='bicubic')
rgb_img_lr = rgb_img_lr.clip(0, 255)
rgb_img_up = scipy.misc.imresize(rgb_img_lr, (shape[0], shape[1]), interp='bicubic') # naive upsampling
# add color jitter
if self.is_train:
rgb_img_up = color_jitter(rgb_img_up, brightness=0.4, contrast=0.4, saturation=0.4, prob=0.5)
rgb_img_up = rgb_img_up.clip(0, 255)
rgb_img_up = np.transpose(rgb_img_up.astype('float32'), (2, 0, 1)) / 255.0
return rgb_img_up
def rgb_processing(self, rgb_img, center, scale, rot, flip, pn):
"""Process rgb image and do augmentation."""
rgb_img = crop(rgb_img, center, scale,
[constants.IMG_RES, constants.IMG_RES], rot=rot)
# flip the image
if flip:
rgb_img = flip_img(rgb_img)
# in the rgb image we add pixel noise in a channel-wise manner
rgb_img[:,:,0] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,0]*pn[0]))
rgb_img[:,:,1] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,1]*pn[1]))
rgb_img[:,:,2] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,2]*pn[2]))
# (3,224,224),float,[0,1]
rgb_img = np.transpose(rgb_img.astype('float32'),(2,0,1))/255.0
return rgb_img
def rgb_processing(self, rgb_img, center, scale, rot, flip, pn):
"""Process rgb image and do augmentation."""
# crop and rotate the image
if self.use_augmentation_rot:
rgb_img = crop(rgb_img, center, scale,
[self.options.img_res, self.options.img_res], rot=rot)
else:
rgb_img = crop(rgb_img, center, scale,
[self.options.img_res, self.options.img_res], rot=0)
# flip the image
if flip:
rgb_img = flip_img(rgb_img)
# in the rgb image we add pixel noise in a channel-wise manner
if self.use_augmentation_rgb:
rgb_img[:,:,0] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,0]*pn[0]))
rgb_img[:,:,1] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,1]*pn[1]))
rgb_img[:,:,2] = np.minimum(255.0, np.maximum(0.0, rgb_img[:,:,2]*pn[2]))
# (3,224,224),float,[0,1]
rgb_img = np.transpose(rgb_img.astype('float32'),(2,0,1))/255.0
return rgb_img