Downloaded SAM 2 annotations are incorrect?

[Applejack]

• Project Type: Downloading instance segmentation annotations
• Operating System & Browser: Windows, Google Chrome
• Dataset link: https://universe.roboflow.com/nirwana/room-segmentation-o7iga/dataset/4

Hi, I am downloading the instance segmentation dataset (not my dataset) from the link above and choosing the annotation format “SAM 2”. However, when decoding the resulting JSON annotations, I see that the annotations are mirrored and rotated compared to the image. Therefore, they are not usable for training as the segmentation masks don’t align with the objects at all. How can I fix this?

Example image from the dataset (this image: Room Segmentation Dataset and Pre-Trained Model by Nirwana), and example result when decoding a RLE mask annotation corresponding to the image:

So you can clearly see that the annotation is wrong. Even though when looking at the image and annotations in Roboflow, the annotations look great.

[iurisilvio]

Hey @Applejack! You’re correct! Our SAM 2 export was generating transposed masks.

It is fixed now. Can you try to generate a new version and export again as SAM 2?

Thank you!