OBB-PBD

Give a UAV/aerial image and a referring expression. The model decodes the oriented box <box> cx cy w h sin2θ cos2θ </box> in a single forward pass (parallel box decoding) and draws it (magenta dashed). On the bundled paper examples the ground truth is drawn too (green solid).

Model: ce-dric/OBB-PBD — LoRA (r=16) on nvidia/LocateAnything-3B · paper: arXiv 2610.35450 · code: ce-dric/OBB-PBD

Aerial image

Referring expression

Prediction (magenta dashed) · GT (green solid)

Paper examples (Figure 2 of arXiv 2610.35450)

Trained on the AVVG subset of RefGeo (UAV imagery); published test-set rotIoU is 0.303 ± 0.012 over three seeds, Acc@0.75 22.0 ± 0.8%. Published results are an average over three training seeds — single-seed rotIoU on an individual sample varies (this demo ships the seed-1 adapter), and in dense scenes the box can land on a neighbouring object. Example images © 2023 S-Lab, RefGeo, S-Lab License 1.0 (non-commercial). Checkpoint and code under the NVIDIA License (non-commercial research use only).