OBB-PBD

Checkpoints for OBB-PBD: Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery (Chang Woo Song, Yeong Jun Koh; Remote Sensing 2026, 18(20), 3535, doi:10.3390/rs18203535; arXiv: 2610.35450).

Code, patch, training configuration and evaluation protocol: github.com/ce-dric/OBB-PBD

Checkpoints

All nine trained runs: OBB-PBD and the two matched controls, three seeds each. Each folder holds the checkpoint chosen on the validation split: LoRA adapters on the language model and the last six MoonViT blocks, plus the fully trained mlp1 projector (adapter_config.json, adapter_model.safetensors, PEFT 0.20.0).

Test results on the complete AVVG test split (6,139 records, 120 images), with 95% image-cluster bootstrap CIs (Table A2 of the paper):

Run	Folder	Selected epoch	rotIoU [95% CI]	aabbIoU [95% CI]
OBB-PBD	obb_pbd/seed1	5	0.3124 [0.2918, 0.3342]	0.3441 [0.3234, 0.3659]
OBB-PBD	obb_pbd/seed2	5	0.3062 [0.2867, 0.3273]	0.3378 [0.3184, 0.3585]
OBB-PBD	obb_pbd/seed3	4	0.2898 [0.2693, 0.3117]	0.3207 [0.3000, 0.3426]
HBB-envelope control	hbb_envelope/seed1	5	0.2527 [0.2364, 0.2702]	0.3465 [0.3261, 0.3682]
HBB-envelope control	hbb_envelope/seed2	5	0.2476 [0.2321, 0.2637]	0.3432 [0.3229, 0.3644]
HBB-envelope control	hbb_envelope/seed3	5	0.2442 [0.2275, 0.2624]	0.3420 [0.3208, 0.3642]
HBB-PBD control	hbb_pbd/seed1	5	0.2523 [0.2358, 0.2696]	0.3552 [0.3337, 0.3780]
HBB-PBD control	hbb_pbd/seed2	5	0.2484 [0.2322, 0.2655]	0.3572 [0.3345, 0.3812]
HBB-PBD control	hbb_pbd/seed3	5	0.2493 [0.2335, 0.2660]	0.3564 [0.3349, 0.3786]

OBB-PBD mean ± standard deviation over the three seeds: rotIoU 0.303 ± 0.012, aabbIoU 0.334 ± 0.012 (zero-shot LocateAnything-3B: 0.1253 / 0.1512).

Usage

The adapters must be loaded on top of LocateAnything-3B at the pinned revision with the backbone patch applied. Without the patch the coordinate slots cannot see the image. Follow the setup in the code repository (download the backbone, apply patches/LocateAnything-3B/combined.diff), then:

hf download ce-dric/OBB-PBD --include "obb_pbd/seed1/*" --local-dir /workspace/hf_ckpt
python scripts/eval_pbd_direct.py --subset avvg --split test --max-eval 0 --device cuda:0 \
  --checkpoint OBB_seed1=/workspace/hf_ckpt/obb_pbd/seed1 --out /workspace/runs/eval_s1

Use --target-mode hbb_envelope for hbb_envelope/* and --target-mode hbb_corner --corner-filter off for hbb_pbd/*. base_model_name_or_path in adapter_config.json points to /workspace/models/LocateAnything-3B, the patched local copy used by the scripts.

License

Distributed under the NVIDIA License (LICENSE): non-commercial research and evaluation use only.

Improved using Qwen. The language model of LocateAnything-3B is based on Qwen2.5-3B-Instruct, licensed under the Qwen RESEARCH LICENSE AGREEMENT (QWEN_RESEARCH_LICENSE.txt, notice in NOTICE). The training data are the AVVG subset of RefGeo (S-Lab License 1.0).

Citation

@article{song2026obbpbd,
  author  = {Song, Chang Woo and Koh, Yeong Jun},
  title   = {{OBB-PBD}: Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery},
  journal = {Remote Sensing},
  volume  = {18},
  number  = {20},
  pages   = {3535},
  year    = {2026},
  doi     = {10.3390/rs18203535},
  eprint  = {2610.35450},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}