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

OBB-PBD. Extending Parallel Box Decoding to Oriented Bounding Boxes for Referring Expression Grounding in Aerial Imagery. By Chang Woo Song and one other, from Chungnam National University. Published on the twenty-sixth of October. Its code repository is worked out: about fifty-five source files, with a readme and a licence.

Referring expression grounding means finding the object a piece of text describes. In overhead imagery, the box around that object has to be oriented rather than horizontal, because aerial objects appear at arbitrary orientations. Most grounding models emit box coordinates one after another. Parallel Box Decoding, or PBD, instead predicts all of its geometry slots in a single forward pass. But it only handles four-coordinate horizontal boxes.

This work extends that block-aligned interface to oriented boxes. The horizontal box's four slots become a six-slot oriented representation, and the angle is encoded with a double-angle parameterization.

The authors do not take a falling loss curve as proof that the extension is working. They verify it directly. Image-shuffle and prompt-shuffle controls, together with visibility and leakage probes, show that the corrected model depends on both the image and the prompt, rather than on the statistics of the prompt alone.

The test is the aerial vehicle visual grounding benchmark, AVVG, whose targets are genuinely oriented. The measure is rotated intersection over union. The zero-shot baseline scores zero point one two five three. The extended model scores zero point three zero three, give or take zero point zero one two, as the mean and standard deviation over three training seeds. That holds across the complete test split of six thousand one hundred and thirty-nine records, spanning all one hundred and twenty images.

Two controls isolate where the gain comes from. Learning oriented boxes, rather than their axis-aligned envelopes, adds zero point zero five five rotated intersection over union beyond a matched control, give or take zero point zero zero eight. The confidence intervals for each seed, at ninety-five percent, all exclude zero. Expanding the block alone yields no gain distinguishable from zero.

The paper's care is in the attribution as much as the score. A model can lower its loss while leaning on prompt statistics, and the shuffle controls rule that out. A gain could also come from the larger block rather than the orientation, and the matched control rules that out too. What remains is improvement traced to learning the orientation itself.