[Question] The processing method on bounding box of VG and RefCOCO are different?

[Richar-Du]

Question

I have seen the issue in #606, and I agree that the processing method of VG is the same as stated in #606 (comment). However, this processing method does not work for RefCOCO, following is an image in RefCOCO processed by expand2square:

And the following is the same image without processing by expand2square:

It seems that the second one is correct.

So I wonder if the processing methods used in RefCOCO and VG are different.

[Richar-Du]

I misunderstand it, sorry for the interrupt.

[NorthSummer]

Hi, I still haven't got it; the normalized [xa, ya, xb, yb] seems correct for VG data after padding, but is not correct for coco data after padding, right?

[boyugou]

Hi, I still haven't got it; the normalized [xa, ya, xb, yb] seems correct for VG data after padding, but is not correct for coco data after padding, right?

Is it true? I also haven't get whether the normalized coordinates in llava665k are adjusted for padding or not

[lzy37ld]

Same