def read_img_caf(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = caffe.io.load_image(fpath) # pixel value range per channel: [0, 1]
img_dat *= 255.
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_PIL(fpath, mean=None):
'''
load image, switch to BGR, subtract mean, and make dims C x H x W for Caffe
'''
img = Image.open(fpath) # pixel value range per channel: [0, 255]
img_dat = np.array(img, dtype=np.float32)
# RGB to BGR
img_dat = img_dat[:, :, ::-1]
# per-channel mean subtraction
if mean is not None:
img_dat -= mean
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
return img_dat
def read_img_cv2(fpath, mean=None):
'''
load image in BGR, subtract mean, and make dims C x H x W for Caffe
'''
img_dat = cv2.imread(fpath) # pixel value range per channel: [0, 255]
# channels already in BGR order
# per-channel mean subtraction
if mean is not None:
img_dat = img_dat.astype(np.float32)
img_dat -= mean
img_dat = img_dat.astype(np.float)
# reorder dimensions
img_dat = mu.hwc_to_chw(img_dat)
# casting to np.float enables plugging into protobuf
return img_dat
