Abstract:Oriented object detection predicts orientation in addition to object location and bounding box. Precisely predicting orientation remains challenging due to angular periodicity, which introduces boundary discontinuity issues and symmetry ambiguities. Inspired by classical works on edge and corner detection, this paper proposes to represent orientation in oriented bounding boxes as a structure tensor. This representation combines the strengths of Gaussian-based methods and angle-coder solutions, providing a simple yet efficient approach that is robust to angular periodicity issues without additional hyperparameters. Extensive evaluations across five datasets demonstrate that the proposed structure tensor representation outperforms previous methods in both fully-supervised and weakly supervised tasks, achieving high precision in angular prediction with minimal computational overhead. Thus, this work establishes structure tensors as a robust and modular alternative for encoding orientation in oriented object detection. We make our code publicly available, allowing for seamless integration into existing object detectors.

What problem does this paper attempt to address?

### What problems does this paper attempt to solve? This paper aims to solve the **challenges of angle prediction in Oriented Object Detection (OOD)**. Specifically, traditional object detection usually uses Horizontal Bounding Boxes (HBB), but some application scenarios (such as aerial imagery and scene text detection) require more precise capture of object orientation information, so Oriented Bounding Boxes (OBB) are introduced. However, there are two main problems when predicting object orientation: 1. **Angular Periodicity**: Due to the periodic nature of angles, a small error at the angle boundary will cause the loss function to increase sharply, even if the predicted value is very close to the true value mathematically. For example, when the predicted angle is close to 0 degrees or 360 degrees, even a small difference between the predicted value and the true value will lead to a large loss. 2. **Symmetry Ambiguities**: For rectangular or square objects, the bounding boxes after rotation by π/2 or -π/2 radians are equivalent, but this will cause the network to wrongly penalize correct predictions during the training process. To solve these problems, the paper proposes a new representation method - **Structure Tensor** to encode the orientation information in the oriented bounding boxes. The structure tensor combines the advantages of the Gaussian method and the angle encoding method, can effectively handle the problems of angular periodicity and symmetry, and does not require additional hyper - parameter adjustment. ### Main contributions 1. **It is the first to propose applying the structure tensor to the orientation representation of oriented bounding boxes**, providing a simple and modular implementation method and making the code public, which is convenient for integration into existing object detectors. 2. **Verified by extensive experiments**, it is proved that the performance of the structure tensor representation method on multiple datasets is better than that of existing methods, especially in terms of angle prediction accuracy, and at the same time, the computational overhead is low. 3. **Combined the advantages of the angle encoding method and the Gaussian - based method**, providing a robust and computationally efficient solution, which elegantly solves the problems of angular periodicity and symmetry. ### Summary This paper solves the two major problems in angle prediction in oriented object detection - angular periodicity and symmetry problems by introducing the structure tensor representation method, thereby improving the detection accuracy and simplifying the model implementation.

Structure Tensor Representation for Robust Oriented Object Detection

Trigonometric-Coded Refined Detector for High Precision Oriented Object Detection

Coupled Dual-Frequency Phase-Shifting Coder for Precise Rotated Angle Representation in Oriented Object Detection

OrientedFormer: An End-to-End Transformer-Based Oriented Object Detector in Remote Sensing Images

ACE: Anchor-Free Corner Evolution for Real-Time Arbitrarily-Oriented Object Detection

On Improving Bounding Box Representations for Oriented Object Detection

Oriented RepPoints for Aerial Object Detection

Highly Efficient Anchor-Free Oriented Small Object Detection for Remote Sensing Images via Periodic Pseudo-Domain

Learning Orientation-Aware Distances for Oriented Object Detection

Spatial Transform Decoupling for Oriented Object Detection

Oriented R-CNN for Object Detection

Projecting Points to Axes: Oriented Object Detection via Point-Axis Representation

Oriented Object Detector with Gaussian Distribution Cost Label Assignment and Task-Decoupled Head

ARS-DETR: Aspect Ratio-Sensitive Detection Transformer for Aerial Oriented Object Detection

RiDOP: A Rotation-Invariant Detector with Simple Oriented Proposals in Remote Sensing Images

Rethinking Boundary Discontinuity Problem for Oriented Object Detection

Linear Gaussian Bounding Box Representation and Ring-Shaped Rotated Convolution for Oriented Object Detection

Oriented objects as pairs of middle lines

Hough Transformed contour Line Segments for object detection

Oriented Object Detection with Transformer

Oriented Bounding Boxes for Small and Freely Rotated Objects