An Oriented Bounding Box (OBB) is a rotated rectangle that tightly encloses an object at an arbitrary angle. Standard axis-aligned bounding boxes have sides parallel to the image edges, which works fine for upright objects but wastes significant area on rotated ones. A ship in satellite imagery, a text line in a tilted document, or an elongated scratch on a metal surface all need rotation-aware boxes to be captured accurately.

An OBB is parameterized as (center_x, center_y, width, height, angle). Detection models must learn to regress the rotation angle on top of the standard four box coordinates. Training uses rotated IoU (SkewIoU) for matching predictions to ground truth, since standard IoU doesn't account for rotation. Evaluation benchmarks like DOTA (Dataset for Object Detection in Aerial Images) test OBB detection on satellite and aerial imagery where objects appear at every angle.

Major detection frameworks now support OBB natively. YOLO26 and YOLOv8-OBB include dedicated OBB prediction heads. Specialized architectures like Oriented R-CNN, RoI Transformer, and LSKNet were built for this task from the ground up. Use cases include aerial/satellite object detection (vehicles, ships, planes), document text detection, industrial defect localization on rotated parts, and agricultural field analysis.