Enhancing Rotated Object Detection via Anisotropic Gaussian Bounding Box and Bhattacharyya Distance

TL;DR

This work tackles rotated object detection by introducing an anisotropic Gaussian bounding box representation and a Bhattacharyya-distance-based loss that is consistent with IoU. By mapping a bounding box ${\mathcal B}(x,y,w,h,\theta)$ to a Gaussian ${\mathcal N}(\mu,\Sigma)$ and applying anisotropic scaling for square-like shapes, the authors define a regression loss ${\mathcal L}_{BD}$ that captures overlap between rotated boxes. They prove key metric properties (non-negativity, identity, symmetry, triangle inequality, scale-invariance) and integrate ${\mathcal L}_{BD}$ into RetinaNet and R3Det, achieving state-of-the-art performance on DOTA-v1.0 and HRSC2016, with notable gains at higher IoU thresholds when using anisotropic bounding boxes. The approach offers a robust, geometry-aware path toward more accurate oriented object localization across diverse aerial and remote-sensing datasets.

Abstract

Detecting rotated objects accurately and efficiently is a significant challenge in computer vision, particularly in applications such as aerial imagery, remote sensing, and autonomous driving. Although traditional object detection frameworks are effective for axis-aligned objects, they often underperform in scenarios involving rotated objects due to their limitations in capturing orientation variations. This paper introduces an improved loss function aimed at enhancing detection accuracy and robustness by leveraging the Gaussian bounding box representation and Bhattacharyya distance. In addition, we advocate for the use of an anisotropic Gaussian representation to address the issues associated with isotropic variance in square-like objects. Our proposed method addresses these challenges by incorporating a rotation-invariant loss function that effectively captures the geometric properties of rotated objects. We integrate this proposed loss function into state-of-the-art deep learning-based rotated object detection detectors, and extensive experiments demonstrated significant improvements in mean Average Precision metrics compared to existing methods. The results highlight the potential of our approach to establish new benchmark in rotated object detection, with implications for a wide range of applications requiring precise and reliable object localization irrespective of orientation.

Figures (10)

• Figure 1: Comparison of Horizontal Bounding Boxes ($x, y, w, h$) and Oriented Bounding Boxes ($x, y, w, h, \theta$)
• Figure 2: The overall pipeline of our proposed method for rotated object detection problem. 'A' indicates the number of categories.
• Figure 3: Relationship between bounding box height and width for various object categories on DOTA-v1.0 dataset. The clustering of data points along the $y = x$ line indicates a significant presence of square-like objects in the dataset.
• Figure 4: Example of isotropic Gaussian case. Both square-like red (ground-truth) and blue (prediction) bounding boxes represent the same Gaussian distribution.
• Figure 5: Comparison of different loss functions over 1000 randomized horizontal bounding boxes pairs. Notably, the Bhattacharyya Loss closely follows the trend of the Complete IoU Loss, indicating their equivalence and similar performance characteristics, whereas GWD Loss and KLD Loss exhibit higher variability and diverge more from the Complete IoU Loss.