search

aminebdj / OpenYOLO3D

Our OpenYOLO3D model achieves state-of-the-art performance in Open Vocabulary 3D Instance Segmentation on ScanNet200 and Replica datasets with up ∼16x speedup compared to the best existing method in literature.
67 stars 4 forks source link

Scaling bounding box of 3D instances for IoU calculation #10

Closed shjung13 closed 2 months ago

shjung13 commented 2 months ago

Hello Authors,

Thank you for the great work!

We are running and visualizing your code and it seems like the bounding boxes of 2D projection of each instance have to be scaled back again using scaling_params.

Line: https://github.com/aminebdj/OpenYOLO3D/blob/1a01f09cfdd5efe356bf48ece991b5dd7aa12589/utils/__init__.py#L197

When we visualize the bboxes, after scaling this, i.e., x_l /= scaling_params[1], I can see their bounding boxes are aligned to the images correctly.

I know this may not affect to selected labels since label map is already scaled, but this seems to affect to normalization.

Would you let me know this finding is correct?

Thanks.

Best, Sanghun

aminebdj commented 2 months ago

Dear @shjung13,

Yes, you are correct, and thanks for letting me know; the bounding box from the projection should be scaled back.

Best,

shjung13 commented 2 months ago

Thank you for the confirmation. I will close this issue.