Possible misalignment when selecting instance features from DINOv2

[YueCheong]

Hi, thanks for the great work!

When selecting features for each object instance, the original 3D points are projected to 640×480 image coordinates, but the processor resizes and center-crops the RGB image to 224×224 before DINOv2 feature extraction (resulting in a 16×16 grid).

Could this cause misalignment when directly using the original (x, y) coordinates to select instance features? Should the same resize and crop be applied to (x, y) before mapping to patch indices?

[YueCheong]

Bug: Potential x/y swap in crop_bbox calculation cause misaligned bounding boxes

When extracting video features ([code],
in process_one_scene, the computed bounding boxes appear misaligned with the objects.

It seems that single_inst_points might be using [y, x] ordering, while later the code assumes [x, y] when saving to crop_bbox.
This mismatch leads to flipped (x0, y0, x1, y1) values and inaccurate visualization outputs.

After swapping x and y, the bounding boxes align correctly with the objects.

[ZzZZCHS]

Are these visualization outputs based on the bbox here? I’ll check this later. If the misalignment is indeed happening, fixing this bug could improve the 2D features. Thanks for pointing this out.