Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence (2106.01883v5)

Abstract: Existing rotated object detectors are mostly inherited from the horizontal detection paradigm, as the latter has evolved into a well-developed area. However, these detectors are difficult to perform prominently in high-precision detection due to the limitation of current regression loss design, especially for objects with large aspect ratios. Taking the perspective that horizontal detection is a special case for rotated object detection, in this paper, we are motivated to change the design of rotation regression loss from induction paradigm to deduction methodology, in terms of the relation between rotation and horizontal detection. We show that one essential challenge is how to modulate the coupled parameters in the rotation regression loss, as such the estimated parameters can influence to each other during the dynamic joint optimization, in an adaptive and synergetic way. Specifically, we first convert the rotated bounding box into a 2-D Gaussian distribution, and then calculate the Kullback-Leibler Divergence (KLD) between the Gaussian distributions as the regression loss. By analyzing the gradient of each parameter, we show that KLD (and its derivatives) can dynamically adjust the parameter gradients according to the characteristics of the object. It will adjust the importance (gradient weight) of the angle parameter according to the aspect ratio. This mechanism can be vital for high-precision detection as a slight angle error would cause a serious accuracy drop for large aspect ratios objects. More importantly, we have proved that KLD is scale invariant. We further show that the KLD loss can be degenerated into the popular $l_{n}$-norm loss for horizontal detection. Experimental results on seven datasets using different detectors show its consistent superiority, and codes are available at https://github.com/yangxue0827/RotationDetection and https://github.com/open-mmlab/mmrotate.

• The paper introduces a novel rotation regression loss by converting rotated bounding boxes into 2-D Gaussian distributions using KLD.
• It dynamically adjusts parameter gradients based on object characteristics to improve detection accuracy for high aspect ratio objects.
• Experimental results on multiple datasets demonstrate state-of-the-art performance and validate the effectiveness of the adaptive loss design.

Overview of "Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence"

Rotated object detection, a crucial component of visual analysis, inherently faces challenges related to high-precision detection, especially for objects with significant aspect ratios. Traditional methods, primarily evolving from horizontal detection paradigms, struggle to maintain performance due to the limitations of current regression loss designs. The paper "Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence" proposes a novel approach utilizing the Kullback-Leibler Divergence (KLD) to enhance detection precision in such contexts.

Core Contribution

The paper introduces a new rotation regression loss framework, transforming the rotated bounding box into a 2-D Gaussian distribution. The KLD between these distributions is used as the regression loss, allowing dynamic adjustment of the parameter weights during optimization. This is particularly beneficial for high-precision detection, as slight inaccuracies in angle estimation can significantly impact objects with large aspect ratios. The authors demonstrate that their proposed approach can dynamically modulate parameter gradients based on object characteristics, a critical aspect for accurate detection.

Methodology

The authors employed a deductive approach to devise a regression loss for rotation detection, treating horizontal detection as a specific instance. By converting rotated bounding boxes to Gaussian distributions, the paper applies KLD to evaluate distribution distance between predicted boxes and ground truths. This strategy ensures adaptability in parameter optimization, helping the model prioritize angle precision when necessary. The paper further validates the scale invariance of KLD, a property not shared by other common loss metrics like Smooth L1 or Gaussian Wasserstein Distance (GWD).

Numerical Results and Claims

Experimental evaluations across seven public datasets and two popular detectors reveal the consistent superiority of the KLD-based approach over traditional methods. Notably, high-precision metrics on challenging scenes, such as those involving large aspect ratio objects, show significant improvement. The paper claims that their method achieves state-of-the-art results, surpassing other techniques in precision and performance stability.

Theoretical and Practical Implications

Theoretically, this research contributes to a better understanding of regression loss design tailored for rotated object detection. By addressing the parameter coupling challenge, the approach offers a unified framework that is adaptable to specific detection tasks. Practically, the improved detection precision has substantial applications in domains like remote sensing and aerial image analysis.

Future Directions

The success of integrating KLD with rotation detection opens pathways for further exploration of statistical measure-based methods in object detection. Future research could explore extending this approach to even more complex shapes like quadrilaterals or irregular objects. Moreover, optimizing computational efficiency while maintaining precision could be a potential area for improvement.

In conclusion, the paper presents a mathematically sound approach to addressing high-precision detection challenges in rotated object detection, underlining the importance of adaptive loss functions in enhancing detection systems.