OBB-PBD teaches LocateAnything-3B to find rotated objects: rotIoU jumps from 0.125 to 0.303

A team led by Chang Woo Song and Yeong Jun Koh at Chungnam National University trained the LocateAnything-3B model to locate objects in aerial imagery from a text description and immediately draw a rotated box around them. On the AVVG benchmark, average rotIoU rose from 0.1253 for the zero-shot baseline to 0.303 ± 0.012 (mean and standard deviation across three training seeds). The work was published in the journal Remote Sensing, volume 18, issue 20, article 3535. The code and all 9 checkpoints are publicly available.

Why a horizontal box doesn't work from a bird's-eye view

In ordinary photographs, objects usually stand more or less vertically, so a horizontal box defined by four coordinates works reasonably well. Drone and satellite imagery is different. A car in a parking lot or at an interchange can be rotated at any angle, and an axis-aligned box ends up capturing a lot of background and neighboring vehicles. The intersection metric drops accordingly.

Most grounding models also predict coordinates sequentially, token by token. Parallel Box Decoding (PBD), built into LocateAnything-3B, works differently: all geometric slots are predicted in a single forward pass. But that block was designed for only four numbers. The authors' task was to fit an angle into this block-based interface without losing speed.

Six slots instead of four

An oriented box is now described by six values: cx, cy, w, h, sin2θ, and cos2θ. The angle is encoded via a doubled argument. Since rotating a rectangle by 180° produces the same shape, directly predicting θ would create a discontinuity at the boundary of the range. The sin2θ/cos2θ pair removes this ambiguity, and the actual angle is later recovered from these two numbers. Decoding still happens in a single pass.

The validation methodology deserves special attention. The authors write that they didn't rely solely on a decreasing loss curve. They ran control tests shuffling images and prompts, along with visibility and leakage checks, to show that the corrected model relies on both the image and the text query — not just on prompt statistics. In a field where it's easy to get a nice-looking loss curve and a useless model, this kind of hygiene matters.

What the numbers showed

• Benchmark: AVVG (aerial vehicle visual grounding), where targets are genuinely rotated at various angles.
• Test set: the full 6,139 records across all 120 images.
• Zero-shot baseline: rotIoU 0.1253.
• OBB-PBD: rotIoU 0.303 ± 0.012 across three seeds.
• Contribution of orientation: training on actual rotated boxes rather than their axis-aligned envelopes adds +0.055 ± 0.008 rotIoU relative to the matched control. 95% confidence intervals for each seed exclude zero.
• Block expansion alone, without orientation training, produces no rotIoU gain distinguishable from zero.
• Box format: (cx, cy, w, h, sin2θ, cos2θ).
• Base model: LocateAnything-3B with Parallel Box Decoding.
• Publication: Remote Sensing 18(20), 3535; code and 9 checkpoints released.

The last point is the most instructive. The gain doesn't come from an architectural add-on but from what the model is actually trained on. Giving it extra slots while still training on horizontal boxes changes nothing.

What to take away

An absolute rotIoU of 0.303 is still far from the ceiling. It's more a working starting point than a ready-made product for territorial monitoring. Still, it has several practical advantages. The approach suits teams working with aerial and satellite imagery: searching for vehicles by description, inventorying parking lots, analyzing traffic. The open checkpoints let you skip training from scratch and instead benchmark against your own baseline on your own data. The sin2θ/cos2θ scheme also transfers to other block-based decoders that need to add an angle without breaking the single-pass interface.

The validation methodology is also worth repeating. The axis-aligned matched control and the shuffling of images and prompts separate the real contribution of orientation from the effect of extra parameters. For any grounding work in non-standard domains, this is a cheap way to verify that the model is actually looking at the image.