About LargeScaleJitter

[ShuaiBai623]

hi, great work! We also try to reimplement the Pix2Seq, we find the absolute coordinate is useful, which is similar to your LargeScaleJitter (pad or crop the image to the fix desired size),
the absolute coordinate means that normalized the position by dividing the fix size.
boxes = boxes / 1333. instead of boxes = boxes / torch.tensor([w, h, w, h], dtype=torch.float32),Then, padding or croppinf the image to the fix desired size is not necessary.

[hanqiu-hq]

Thanks.
We also use absolute coordinate as described in Pix2Seq.

Pretrained-Pix2Seq/playground/pix2seq/pix2seq.py

Lines 88 to 90 in 7d908d4

	scaled_box = box * scale_factor
	scaled_box = box_cxcywh_to_xyxy(scaled_box)
	box_tokens = (scaled_box / 1333 * self.num_bins).floor().long().clamp(min=0, max=self.num_bins)

The conversion of relative coordinate in transforms.py is because we use the same dataloader as DETR.
And for the large Scale jittering, we basically follow the same pipeline proposed in CopyPaste which is cited in Pix2seq.

[baiyongrui]

Hello, does using absolute coordinate gets a better AP?

[seanzhuh]

The only difference between relative and absolute coordinate is the normalization factor ? Absolute will normalize by the longest image size instead of the actual image size, which relative coordinate does, am I right ?