Bboxes Bias And Meaning of Crop

[ronyyuan]

Hi,
Great work on the dataset! I have a question about using the MultiID-2M dataset that I'm hoping you can help clarify.
I noticed that when the 'crop' attribute doesn't start from [0, 0], all the face bounding boxes in the 'bboxes' field seem to have an offset equal to the starting coordinates of the 'crop' attribute.
Could you please explain what the 'crop' attribute represents?
Specifically, to get the true face bounding box coordinates in the original image, should I be applying this transformation:

[bboxes[0] - crop[0], bboxes[1] - crop[1], bboxes[2] - crop[0], bboxes[3] - crop[1]]

For more clearly, please ref to the following visualization

The right bounding box is draw by the raw 'bboxes', while the left by the mentioned transformation.

Looking forward to your clarification.
Best regards,
Rony

[EngDoge]

the crop maybe correspond to the original image coordinates, the "url" in the json file

[Doby-Xu]

Apologies for the confusion earlier. In short, your proposed remedy is correct.

The term "Crop" refers to cropping out the text regions from the raw image, which is necessary if you download the image directly from the provided URL. Initially, we only intended to open-source the URLs.

However, the bounding boxes (bboxes) are also defined with respect to the raw image. With your transformation:

[bboxes[0] - crop[0], bboxes[1] - crop[1], bboxes[2] - crop[0], bboxes[3] - crop[1]]

it should indeed work as expected.

We will update this clarification in the Multi-ID 2M README soon.

In your case, it seems the bounding box visualization is misaligned — the y1 coordinates are 613 and 173, which indicates a large vertical offset.

[ronyyuan]

Thank you very much. I fully understand now.