merge before #1506

A question here, in 0.x version, the pad ratio was set as 0.4 and 0.2 by default. I am not sure if this matters to the test performance. Though personally, I think it should be set to 0 as this PR, maybe someone should test the model to see if it works as expected.

A question here, in 0.x version, the pad ratio was set as 0.4 and 0.2 by default. I am not sure if this matters to the test performance. Though personally, I think it should be set to 0 as this PR, maybe someone should test the model to see if it works as expected.

mmocr/mmocr/datasets/pipelines/crop.py

Lines 87 to 97 in 26bc471

	def crop_img(src_img, box, long_edge_pad_ratio=0.4, short_edge_pad_ratio=0.2):
	"""Crop text region with their bounding box.
	Args:
	src_img (np.array): The original image.
	box (list[float | int]): Points of quadrangle.
	long_edge_pad_ratio (float): Box pad ratio for long edge
	corresponding to font size.
	short_edge_pad_ratio (float): Box pad ratio for short edge
	corresponding to font size.
	"""

I have tested. Difference image shape will affect the performance.
In the 0.x version, the crop image is only used for ocr.py, because of bad performance of text detection

A question here, in 0.x version, the pad ratio was set as 0.4 and 0.2 by default. I am not sure if this matters to the test performance. Though personally, I think it should be set to 0 as this PR, maybe someone should test the model to see if it works as expected.

mmocr/mmocr/datasets/pipelines/crop.py

Lines 87 to 97 in 26bc471

	def crop_img(src_img, box, long_edge_pad_ratio=0.4, short_edge_pad_ratio=0.2):
	"""Crop text region with their bounding box.
	Args:
	src_img (np.array): The original image.
	box (list[float | int]): Points of quadrangle.
	long_edge_pad_ratio (float): Box pad ratio for long edge
	corresponding to font size.
	short_edge_pad_ratio (float): Box pad ratio for short edge
	corresponding to font size.
	"""

I have tested. Difference image shape will affect the performance. In the 0.x version, the crop image is only used for ocr.py, because of bad performance of text detection

Ok. It is noteworthy that some converters such as totaltext converter also used the crop_image and the default setting. I am not sure if it affects totaltext's performance.

A question here, in 0.x version, the pad ratio was set as 0.4 and 0.2 by default. I am not sure if this matters to the test performance. Though personally, I think it should be set to 0 as this PR, maybe someone should test the model to see if it works as expected.

mmocr/mmocr/datasets/pipelines/crop.py

Lines 87 to 97 in 26bc471

	def crop_img(src_img, box, long_edge_pad_ratio=0.4, short_edge_pad_ratio=0.2):
	"""Crop text region with their bounding box.
	Args:
	src_img (np.array): The original image.
	box (list[float | int]): Points of quadrangle.
	long_edge_pad_ratio (float): Box pad ratio for long edge
	corresponding to font size.
	short_edge_pad_ratio (float): Box pad ratio for short edge
	corresponding to font size.
	"""

I have tested. Difference image shape will affect the performance. In the 0.x version, the crop image is only used for ocr.py, because of bad performance of text detection

Ok. It is noteworthy that some converters such as totaltext converter also used the crop_image and the default setting. I am not sure if it affects totaltext's performance.

mmocr/tools/data/textrecog/totaltext_converter.py

Line 302 in 26bc471

dst_img = crop_img(image, anno['bbox'])

Maybe it's a bug. I check other data_converters, and they are all crop_img(image, anno['box], 0, 0)

@Harold-lkk Need a followup PR to avoid BC-breaking changes