Capsar/DL-GenISP

Mismatch of the image size between rawpy's output and bbox annotations

QiuJueqin opened this issue · 2 comments

QiuJueqin commented

Hi Capsar,

For Nikon 750 dataset, the images read by rawpy is 2640*3968 (without any postprocess), but in annotation files (e.g., https://github.com/igor-morawski/RAW-NOD/blob/main/annotations/Nikon/raw_new_Nikon750_train.json), the image size is 2624*3936 ("width": 3936, "height": 2624).

How did you handle this mismatch? Thanks.

Capsar commented

Hi, I had not noticed this difference.

However, in confirming the boundingboxes visually via the methods in: visualize_single_image.py no noticeable difference can be seen. Indicating that the size of the images is correct.

Where exactly are you encountering an issue?

Kind regards,
Caspar

QiuJueqin commented

Hi Capsar, thanks for your reply.

I meant the ground-truth annotations mismatched from the images, but your visualize_single_image.py visualizes the predicted annotations.

Try to visualize image DSC01803 in Sony dataset with the ground-truth bboxes, and you can notice obvious misalignment of the bboxes from the objects-of-interest in the image. Please note that the image to be plotted on should be processed by RawPy or any other RAW editting software, with lens distortion correction disabled, rather than being directly extracted from jpg thumbnail embedded in raw file.