OBB-PBD: Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery

Source abstract

Referring expression grounding in overhead imagery requires oriented bounding boxes rather than horizontal ones, since aerial objects appear at arbitrary orientations. Most grounding models emit box coordinates sequentially; Parallel Box Decoding (PBD) instead predicts all geometry slots in one forward pass, but only for four-coordinate horizontal boxes. We extend this block-aligned interface to a six-slot oriented representation using a double-angle parameterization, and verify the extension with image/prompt-shuffle controls and visibility/leakage probes rather than assuming correctness from a decreasing loss curve; these confirm the corrected model depends on both the image and the prompt rather than on prompt statistics alone. On AVVG, an aerial vehicle visual-grounding benchmark with genuinely oriented targets, the extended model improves rotated IoU by 149.4% over the zero-shot baseline (0.1253 to 0.3124) across the complete 6,139-record test split spanning all 120 images. Learning orientation adds 0.0597 rotIoU beyond a matched control trained without orientation (95% CI excluding zero), whereas block expansion alone yields no rotIoU gain distinguishable from zero.