PaperlayerView exact revision on arXiv

Source-linked AI summary

Learning Modulated Loss for Rotated Object Detection

Wen Qian, Xue Yang, Silong Peng, Yue Guo, Junchi Yan

arXiv:1911.08299v3cs.CV

TL;DR

Five-parameter rotated-box regression can suffer discontinuous losses and inconsistent parameter scales, producing rotation sensitivity error. The paper addresses this with modulated rotation losses and eight-parameter regression in RSDet, which achieves state-of-the-art DOTA performance and generalizes across datasets and detectors.

  • Problem

    Five-parameter rotated-box regression combines angle periodicity, width-height exchange, and differently scaled parameters, creating rotation sensitivity error associated with unstable learning and degraded detection performance.

  • Method

    RSDet combines modulated rotation loss with eight-parameter regression to smooth loss discontinuities and address inconsistent parameter regression.

  • Results

    RSDet achieves state-of-the-art performance on DOTA and demonstrates generalization and robustness across different datasets and multiple detectors.

  • Takeaways & Limitations

    The proposed modulated losses and eight-parameter regression provide an end-to-end approach for handling rotation sensitivity error in rotated object detection.

Abstract

from arXiv · show

Popular rotated detection methods usually use five parameters (coordinates of the central point, width, height, and rotation angle) to describe the rotated bounding box and l1-loss as the loss function. In this paper, we argue that the aforementioned integration can cause training instability and performance degeneration, due to the loss discontinuity resulted from the inherent periodicity of angles and the associated sudden exchange of width and height. This problem is further pronounced given the regression inconsistency among five parameters with different measurement units. We refer to the above issues as rotation sensitivity error (RSE) and propose a modulated rotation loss to dismiss the loss discontinuity. Our new loss is combined with the eight-parameter regression to further solve the problem of inconsistent parameter regression. Experiments show the state-of-art performances of our method on the public aerial image benchmark DOTA and UCAS-AOD. Its generalization abilities are also verified on ICDAR2015, HRSC2016, and FDDB. Qualitative improvements can be seen in Fig 1, and the source code will be released with the publication of the paper.

1. Introduction

Rotated detection is important for densely arranged or elongated objects, but five-parameter representations introduce rotation sensitivity errors that can destabilize training and degrade detection. The paper formulates these errors and proposes modulated losses with eight-parameter regression in RSDet.

  • Rotated detection supports more accurate localization than horizontal detection in scene text, aerial imagery, face detection, and license plate detection.
  • Five-parameter rotation detection suffers loss discontinuity when angle periodicity causes large parameter changes after small physical rotations.The angle can jump at its range boundary, while width and height may exchange roles.
  • Different measurement units and IoU relationships among angle, width, height, and center coordinates create inconsistent five-parameter regression.The paper identifies this inconsistency as a second component of rotation sensitivity error.
  • Rotation sensitivity error can cause training instability and detection performance degeneration.
  • The paper proposes modulated rotation losses for five- and eight-parameter systems to smooth discontinuities and introduces the end-to-end RSDet detector.The five-parameter loss is termed ℓ5p_mr, while the eight-parameter treatment uses ℓ8p_mr.

2. Related Work

Rotated detection extends conventional horizontal detectors for scenes requiring precise localization, but existing approaches had not addressed the inherent loss discontinuity in rotated-box parameterizations. The paper motivates this problem through five-parameter angle and width-height behavior.

  • A near-horizontal rotation example yields a large ℓ1 loss because reference, ground-truth, and predicted boxes receive sharply different five-parameter descriptions.
  • Horizontal detectors generate boxes only along the horizontal direction, limiting their applicability to densely arranged or elongated objects in scene text and aerial images.
  • Rotated object detectors use rotated bounding boxes to localize objects more accurately than horizontal boxes in scene text and aerial imagery.
  • OpenCV's five-parameter definition exchanges width and height at a rotation boundary and restricts the angle parameter θ to -90° through 0°.
  • Existing rotated detectors described in the related work did not consider inherent loss discontinuity, which the paper links to learning stability and final detection performance.

3. Proposed Method

The method identifies rotation sensitivity error from boundary discontinuities and inconsistent five-parameter regression, then combines modulated rotation losses with eight-parameter corner regression. This design targets smoother training and improved detection performance.

  • Eight-Parameter Regression: Eight-parameter regression represents rotated boxes by four clockwise corner coordinates, providing consistent parameter units and supporting quadrilateral descriptions.Corner ordering is performed before regression to avoid large losses when the predicted pose is correct.
  • Rotation Sensitivity Error: Rotation sensitivity error combines angle-boundary discontinuity, height–width exchange, and inconsistent measurement-unit regression in five-parameter rotated detection.These issues can cause training instability and detection performance degeneration.
  • Rotation Sensitivity Error: Five-parameter regression maps angle, width, height, and center-point changes to IoU through different relationship shapes, making their joint regression inconsistent.The paper describes angle–IoU as polynomial, width/height–IoU as combined linear and inverse-proportion, and center-point–IoU as symmetric linear.
  • Five-Parameter Modulated Rotation Loss: The five-parameter modulated rotation loss uses a correction term that removes angular periodicity and width–height exchange effects, then takes the minimum with ℓ1-loss.The resulting ℓmr curve is continuous, including at the angle-boundary mutation.
  • Eight-Parameter Modulated Rotation Loss: Eight-parameter regression still has boundary discontinuity because corner sequences cyclically shift after slight rotations, so the loss evaluates clockwise, unchanged, and counterclockwise vertex orders.The eight-parameter modulated loss takes the minimum across these three orderings.
  • Overall Design: RSDet combines eight-parameter regression with modulated rotation loss in an end-to-end detector and reports more stable training and state-of-the-art benchmark performance.The paper also reports generalization and robustness across different datasets and detectors.

4. Experiments

Experiments evaluate the proposed techniques through ablations, training-stability analyses, comparisons on DOTA and UCAS-AOD, and tests on other datasets and detector architectures. The results report improved accuracy, more stable convergence, and generalization across settings.

  • Ablation Study: The ablation study on DOTA measures modulated rotation loss and eight-parameter regression using RetinaNet-H as the baseline.The experiments also examine backbones, augmentation, data balance, and regression refinement.
  • Ablation Study: 2.35% mAP is gained by replacing smooth-ℓ1 with ℓmr, while eight-parameter regression adds 3.45% mAP; combining both reaches 66.77% mAP.The reported gains come from the DOTA ablation experiments.
  • Training Stability: Training converges more stably after using modulated rotation losses, as shown by comparisons of training loss curves.The stability analysis compares models using different rotation-loss formulations.
  • Backbone, Data Augmentation, and Data Balance: Data augmentation raises baseline DOTA performance by 4.22% to 70.79%, while ResNet50, ResNet101, and ResNet152 reach 71.22%, 72.16%, and 73.51%, respectively.The augmentation operations include horizontal and vertical flipping, graying, and random rotation; copying underrepresented categories adds 0.43%.
  • Other Detectors and Datasets: The techniques improve a rotating Faster RCNN baseline by 1.6% and 2.84%, and validation experiments report positive results on ICDAR2015 and HRSC2016.The paper describes the detector as competitive on scene-text data and comparable to state of the art on HRSC2016.
  • Overall Evaluation: On DOTA, RSDet’s overall mAP is 1.3% higher than the existing best method, while UCAS-AOD reaches 96.50% for the OBB task.The DOTA comparison uses official evaluation-server results without multi-scale testing, oversized backbones, or model integration.

5. Conclusion

The paper identifies rotation sensitivity error in region-based rotated detectors and addresses its two components with modulated rotation loss and eight-parameter regression. The resulting RSDet is trained end-to-end and reports state-of-the-art DOTA performance with generalization and robustness across datasets and detectors.

  • 5. Conclusion: Rotation sensitivity error comprises loss discontinuity and five-parameter regression inconsistency in region-based rotated object detectors.The paper formally identifies and formulates RSE as the central problem addressed by the method.
  • 5. Conclusion: RSDet combines modulated rotation loss ℓmr with eight-parameter regression to address loss discontinuity and regression inconsistency in an end-to-end detector.The two techniques target the two stated components of RSE.
  • 5. Conclusion: Extensive experiments report state-of-the-art performance on DOTA and good generalization and robustness across different datasets and multiple detectors.The conclusion summarizes the reported evaluation scope without specifying individual benchmark values.
Loading 1911.08299v3…