Oriented Object Detection from Horizontal Box Supervision

This direction is merged into another direction: oriented object detection. Its last version stays readable and citable here.

Version 1 of 19 papers17 conceptsfrozen 2026-10-04

In computer vision, detecting objects with precise orientation—like ships in a harbor or planes on a runway—usually requires expensive "oriented" bounding box (OBB) annotations. Unlike standard horizontal boxes (HBox), which only require two points (top-left and bottom-right), oriented boxes require four corners or a specific rotation angle. This makes the annotation process significantly more labor-intensive.

A growing body of research demonstrates that we can bypass this cost. By leveraging geometric constraints and object symmetries, researchers have developed frameworks that train high-performance oriented detectors using only standard, cheap horizontal box labels.

The Circumscribed Box Constraint

The fundamental challenge of training an oriented detector with horizontal labels is the "supervision gap": the model predicts a rotated rectangle, but the ground truth is an axis-aligned one.

To bridge this gap, researchers like Xue Yang and Xuehui Wang in H2RBox propose a projection mechanism. During training, the model predicts an oriented box. This box is then mathematically converted into its minimal axis-aligned circumscribed rectangle—the tightest horizontal box that can contain the rotated prediction.

The loss is then calculated between this projected horizontal box and the actual horizontal ground truth. This provides a gradient signal: if the projected box is too wide or too tall compared to the ground truth, the model adjusts the predicted OBB's dimensions and angle to fit.

Resolving Rotational Ambiguity

A single horizontal box can circumscribe many different rotated boxes. For example, a square HBox could contain a diamond shape or a smaller square rotated at various angles. This creates an ambiguity problem: how does the model know which specific orientation is correct?

Researchers have introduced three primary strategies to resolve this:

  1. Geometric Consistency: In H2RBox, the authors enforce consistency across different views. If an image is rotated, the model's prediction for the object should rotate by the same amount. By enforcing that the predicted OBB remains "stable" under transformation, the model learns to find the true orientation that satisfies the horizontal constraints in all views.
  2. Reflection Symmetry: Many objects, such as ships and aircraft, possess inherent reflection symmetry. Yi Yu and Yue Zhou expanded on this in H2RBox-v2. They introduced a self-supervised branch that leverages this symmetry, allowing the model to refine its angle predictions by ensuring the object looks "correct" when flipped or rotated.
  3. Minimal Cues: Sometimes, a tiny amount of extra information is enough. Javed Iqbal and colleagues showed in their OAOD framework that providing just a single orientation vector (a simple line indicating direction) alongside an HBox is sufficient to achieve higher accuracy than complex four-corner supervision.

Knowledge Transfer and Domain Adaptation

Another approach involves "teaching" a model what rotation looks like in one context and applying it to another. Tianyu Zhu and Hamid Rezatofighi developed the Knowledge Combination for Rotated detection (KCR) framework.

In KCR, a model is co-trained on a source dataset that has rotated labels and a target dataset that only has horizontal labels. By projecting predicted boxes into their horizontal equivalents for the target dataset, the model transfers the general concept of "object-ness and rotation" from the labeled source to the weakly-labeled target.

Why It Works

This research direction works because the geometry of a horizontal bounding box actually contains a surprising amount of information about the object inside it. As an object rotates, its horizontal footprint changes in a predictable way.

By combining the circumscribed box constraint with rotational data augmentation, these models can observe how the horizontal footprint "flickers" or shifts as the image is transformed. This allows the model to mathematically "triangulate" the true orientation of the object. As shown by H2RBox-v2, these weakly-supervised methods can now achieve performance parity with fully supervised models, effectively making expensive oriented annotations unnecessary for many industrial and aerial imaging applications.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_oriented_object_detection_from_horizonta_53eb305b,
  title = {Oriented Object Detection from Horizontal Box Supervision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/oriented-object-detection-from-horizontal-box-supervision/rdr_53eb305b4dad45128e2707bef77bc302},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_oriented_object_detection_from_horizonta_53eb305b_v1,
  title = {Oriented Object Detection from Horizontal Box Supervision},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/oriented-object-detection-from-horizontal-box-supervision/rdr_53eb305b4dad45128e2707bef77bc302/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}