RSDet++: Point-based Modulated Loss for More Accurate Rotated Object Detection (2109.11906v1)

Abstract: We classify the discontinuity of loss in both five-param and eight-param rotated object detection methods as rotation sensitivity error (RSE) which will result in performance degeneration. We introduce a novel modulated rotation loss to alleviate the problem and propose a rotation sensitivity detection network (RSDet) which is consists of an eight-param single-stage rotated object detector and the modulated rotation loss. Our proposed RSDet has several advantages: 1) it reformulates the rotated object detection problem as predicting the corners of objects while most previous methods employ a five-para-based regression method with different measurement units. 2) modulated rotation loss achieves consistent improvement on both five-param and eight-param rotated object detection methods by solving the discontinuity of loss. To further improve the accuracy of our method on objects smaller than 10 pixels, we introduce a novel RSDet++ which is consists of a point-based anchor-free rotated object detector and a modulated rotation loss. Extensive experiments demonstrate the effectiveness of both RSDet and RSDet++, which achieve competitive results on rotated object detection in the challenging benchmarks DOTA1.0, DOTA1.5, and DOTA2.0. We hope the proposed method can provide a new perspective for designing algorithms to solve rotated object detection and pay more attention to tiny objects. The codes and models are available at: https://github.com/yangxue0827/RotationDetection.

• The paper identifies the Rotation Sensitivity Error (RSE) in existing detection methods and proposes a modulated rotation loss to smooth training and improve results.
• It introduces a point-based, anchor-free RSDet++ framework that uses eight-parameter regression for superior detection of rotated objects, especially small ones.
• Extensive evaluations on DOTA benchmarks show RSDet++ achieves over 2% mAP improvement and enhanced stability compared to baseline models.

Overview of RSDet++: Point-based Modulated Loss for More Accurate Rotated Object Detection

The paper presents RSDet++, a sophisticated approach to enhance the accuracy of rotated object detection. This paper addresses the critical issue of loss discontinuity, referred to as Rotation Sensitivity Error (RSE), which occurs in both five-parameter and eight-parameter object detection methods. The authors introduce a novel modulated rotation loss to mitigate this problem, alongside a point-based anchor-free detection framework aimed at improving performance on small objects. The method is evaluated on several challenging benchmarks, including DOTA1.0, DOTA1.5, and DOTA2.0.

Key Contributions

1. Rotation Sensitivity Error Identification: The paper classifies the discontinuity of loss in established five-parameter and eight-parameter detection methods as RSE. This phenomenon can lead to significant performance degradation due to instability during training.
2. Modulated Rotation Loss: A modulated rotation loss is introduced, which incorporates boundary constraints, enabling smoother and more consistent loss curves during training. This addresses the angular periodicity and unit inconsistency in parameter regression, which are prevalent challenges in traditional methods.
3. RSDet++ Framework: The RSDet++ framework is proposed, utilizing a point-based approach for bounding box regression. This model employs eight parameters for detecting rotated objects, which mitigates issues inherent in anchor-based methods—particularly the challenges posed by small objects.
4. Extensive Evaluation: Extensive experiments were conducted on various benchmarks, demonstrating the superior performance of RSDet and RSDet++ in both stability during training and final detection results.

Numerical Results

The RSDet framework achieves a significant mAP improvement, outperforming baseline models on the DOTA benchmark by more than 2%, with RSDet++ further enhancing detection accuracy. Notably, RSDet++ attains competitive results on DOTA1.5 and DOTA2.0, datasets that emphasize detecting tiny objects.

Implications and Future Work

Practically, this research provides a robust method for enhancing real-world applications such as aerial surveillance and automated navigation systems where rotated objects are common. Theoretically, it highlights the importance of addressing loss discontinuity in deep learning models. Looking forward, the approach could be extended to other domains of object detection, potentially integrating more advanced deep learning architectures to further boost performance.

In conclusion, the RSDet++ framework and its underlying loss function offer considerable advancements in object detection methodologies. By resolving key issues in existing systems, this work contributes to more precise detection capabilities in complicated visual environments and sets a foundation for future explorations in rotated object detection frameworks.