def map_fn(image):
input_processor = dataloader.DetectionInputProcessor(
image, image_size)
input_processor.normalize_image(mean_rgb, stddev_rgb)
input_processor.set_scale_factors_to_output_size()
image = input_processor.resize_and_crop_image()
image_scale = input_processor.image_scale_to_original
return image, image_scale
def image_preprocess(image, image_size, mean_rgb, stddev_rgb):
"""Preprocess image for inference.
Args:
image: input image, can be a tensor or a numpy arary.
image_size: single integer of image size for square image or tuple of two
integers, in the format of (image_height, image_width).
mean_rgb: Mean value of RGB, can be a list of float or a float value.
stddev_rgb: Standard deviation of RGB, can be a list of float or a float
value.
Returns:
(image, scale): a tuple of processed image and its scale.
"""
input_processor = dataloader.DetectionInputProcessor(image, image_size)
input_processor.normalize_image(mean_rgb, stddev_rgb)
input_processor.set_scale_factors_to_output_size()
image = input_processor.resize_and_crop_image()
image_scale = input_processor.image_scale_to_original
return image, image_scale
def build_preprocess(self, image):
config = self.config
img_size = config.max_size
bgr = True # cv2 load image is bgr
p_image = image
if bgr:
# to RGB, efficientdet is trained with PIL
p_image = p_image[:, :, ::-1]
#input_processor = dataloader.DetectionInputProcessor(p_image, img_size)
input_processor = dataloader.DetectionInputProcessor(
p_image, img_size, config.short_edge_size)
# make image [0,1] and -mean/var
input_processor.normalize_image()
input_processor.set_scale_factors_to_output_size()
# here the original efficientdet pad image to (max_size, max_size)
p_image = input_processor.resize_and_crop_image()
p_image_scale = input_processor.image_scale_to_original
p_image = tf.expand_dims(p_image, 0) # [1, H, W, C]
return p_image, p_image_scale