About SCALE and MAX_SIZE

[lilichu]

All images of my dataset is 720 * 1280 and 1280 * 720(height*weight), in yaml:

  SCALES: (800,)
  MAX_SIZE: 1333

Does it mean the image will be resize as 800 * 1280 and 1280 * 800?

[ppwwyyxx]

No it will be resized to 750x1333. See

Detectron/detectron/utils/blob.py

Lines 100 to 125 in b5dcc0f

	def prep_im_for_blob(im, pixel_means, target_size, max_size):
	"""Prepare an image for use as a network input blob. Specially:
	- Subtract per-channel pixel mean
	- Convert to float32
	- Rescale to each of the specified target size (capped at max_size)
	Returns a list of transformed images, one for each target size. Also returns
	the scale factors that were used to compute each returned image.
	"""
	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)
	# Prevent the biggest axis from being more than max_size
	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
	)
	return im, im_scale

for the implementation.

[chelixuan]

Hi! The ratio of length to width of the picture in Ms COCO is different, if we re-scale the images that their shorter side is 800, we get the different ratio. In the end, the longer side of images is different. The size of the training images is different . But the input_image should be at the same size. How can I solve this problem? Thx~ @ppwwyyxx @

[ppwwyyxx]

Detectron/detectron/utils/blob.py

Lines 67 to 97 in dfbad58

	def im_list_to_blob(ims):
	"""Convert a list of images into a network input. Assumes images were
	prepared using prep_im_for_blob or equivalent: i.e.
	- BGR channel order
	- pixel means subtracted
	- resized to the desired input size
	- float32 numpy ndarray format
	Output is a 4D HCHW tensor of the images concatenated along axis 0 with
	shape.
	"""
	if not isinstance(ims, list):
	ims = [ims]
	max_shape = np.array([im.shape for im in ims]).max(axis=0)
	# Pad the image so they can be divisible by a stride
	if cfg.FPN.FPN_ON:
	stride = float(cfg.FPN.COARSEST_STRIDE)
	max_shape[0] = int(np.ceil(max_shape[0] / stride) * stride)
	max_shape[1] = int(np.ceil(max_shape[1] / stride) * stride)
	num_images = len(ims)
	blob = np.zeros(
	(num_images, max_shape[0], max_shape[1], 3), dtype=np.float32
	)
	for i in range(num_images):
	im = ims[i]
	blob[i, 0:im.shape[0], 0:im.shape[1], :] = im
	# Move channels (axis 3) to axis 1
	# Axis order will become: (batch elem, channel, height, width)
	channel_swap = (0, 3, 1, 2)
	blob = blob.transpose(channel_swap)
	return blob