OBB-PBD extends parallel box decoding to oriented boxes, lifting aerial grounding rotIoU from 0.13 to 0.30
ChungnamNationalUniversity · hf · 2026-10-27
Chungnam National University's OBB-PBD extends Parallel Box Decoding to oriented bounding boxes for referring expression grounding in aerial imagery. Adding a LoRA to LocateAnything-3B, one forward pass emits six tokens in parallel — cx, cy, w, h plus sin2θ/cos2θ double-angle parameterization — to predict oriented boxes, validated with control probes. On RefGeo (AVVG) the zero-shot baseline scores mean rotIoU 0.1253 vs 0.303 for OBB-PBD, of which +0.055 comes from learning orientation. Published in Remote Sensing 18(20), arXiv:2610.35450.
More from Research
- Screening 307 AI-for-science papers in one week: the test that matters is out-of-distribution — bravo_abad · 2026-10-28
- Why K=4 beats K=32 in group-sampled RL training: filtering easy prompts faster — gowthami_s · 2026-10-28
- RL trick: keep sampling all-wrong groups so rare positives get advantage ~1, not 0.75 — gowthami_s · 2026-10-28
- What 'hard prompts' look like in vision: rare concept combos like armadillos in berets — gowthami_s · 2026-10-28
- Robotics may need to go beyond language: demos as a test-time steering modality — Ed__Johns · 2026-10-28
- VybBench: open benchmark scores image models' design fidelity with numbers, not votes — thetripathi58 · 2026-10-28