Feed
OBB-PBD: Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery
Teaching a vision-language model to draw tilted boxes around vehicles in drone photos, all at once.
0.303
rotIoU
Rotated box overlap on AVVG
More than doubles the untuned model on all 6139 test records
How it works
Tilted cars, straight boxes
Drone images show vehicles at any angle. Parallel Box Decoding predicts boxes in one pass, but only upright four-number boxes.
Six slots, one pass
Add center, size and sin/cos of twice the angle as six slots. Probes then catch a bug hiding the image from them.
Better tilted boxes
Rotated overlap rises from 0.1253 to 0.303. Learning orientation alone adds 0.055; extra slots alone add nothing.
From the paper