OBB-PBD extends parallel box decoding to oriented aerial boxes
Original title: OBB-PBD: Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery
Reported rotIoU on the full 6139-record AVVG test set rises from 0.1253 zero-shot to 0.303.
Most grounding models emit box coordinates sequentially, while Parallel Box Decoding predicts horizontal box slots in one pass. OBB-PBD changes that interface to six slots with a double-angle parameterization for oriented aerial objects. Shuffle and leakage probes indicate the corrected model uses both the image and the prompt, not prompt statistics alone. Oriented-box learning adds 0.055 rotIoU over a matched control; expanding the block alone does not.
Unconfirmed: Reported metrics come from the authors' AVVG runs and are not independently replicated here.
- The AVVG test split covers 6139 records and all 120 images.
- Mean rotIoU is 0.303 with standard deviation 0.012 over three seeds.
- The oriented-box gain is 0.055 plus or minus 0.008 rotIoU.
- Per-seed 95% confidence intervals for that gain exclude zero.
- Visibility and leakage probes were used instead of relying on a falling loss.
From the original article
1 min read at huggingface.co
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.
By Chang Woo Song, Yeong Jun Koh · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.