PaperlayerView exact revision on arXiv

Source-linked AI summary

On the Arbitrary-Oriented Object Detection: Classification based Approaches Revisited

Xue Yang, Junchi Yan

arXiv:2003.05597v4cs.CVcs.AI

TL;DR

Existing regression-based rotation detectors face boundary problems because ideal angle predictions can fall outside defined ranges. This paper reformulates angle prediction as classification, introduces Circular Smooth Label and Densely Coded Label techniques, and reports state-of-the-art performance on HRSC2016 and DOTA-v1.0.

  • Problem

    Regression-based rotation detectors suffer boundary problems because ideal predictions can lie beyond the defined range, increasing boundary loss and complicating learning.

  • Method

    The paper transforms angle prediction into classification, uses Circular Smooth Label to handle periodicity and adjacent-angle tolerance, and introduces Densely Coded Label to reduce encoding length.

  • Results

    The detector achieves state-of-the-art performance on HRSC2016 and DOTA-v1.0, while CSL and DCL are effective as independent components across different detectors.

  • Takeaways & Limitations

    The classification-based baseline addresses the boundary problem by restricting angle predictions and accommodates circular angles with little accuracy error.

  • Takeaways & Limitations

    The paper identifies labeling workload and detection limitations for existing instance-segmentation approaches in aerial images, including background areas, dense-scene suppression, and channel-assignment constraints.

Abstract

from arXiv · show

Arbitrary-oriented object detection has been a building block for rotation sensitive tasks. We first show that the boundary problem suffered in existing dominant regression-based rotation detectors, is caused by angular periodicity or corner ordering, according to the parameterization protocol. We also show that the root cause is that the ideal predictions can be out of the defined range. Accordingly, we transform the angular prediction task from a regression problem to a classification one. For the resulting circularly distributed angle classification problem, we first devise a Circular Smooth Label technique to handle the periodicity of angle and increase the error tolerance to adjacent angles. To reduce the excessive model parameters by Circular Smooth Label, we further design a Densely Coded Labels, which greatly reduces the length of the encoding. Finally, we further develop an object heading detection module, which can be useful when the exact heading orientation information is needed e.g. for ship and plane heading detection. We release our OHD-SJTU dataset and OHDet detector for heading detection. Extensive experimental results on three large-scale public datasets for aerial images i.e. DOTA, HRSC2016, OHD-SJTU, and face dataset FDDB, as well as scene text dataset ICDAR2015 and MLT, show the effectiveness of our approach.

1 INTRODUCTION

Existing rotation detectors commonly regress angles or quadrilateral corners, but boundary discontinuities arise when ideal predictions fall outside defined ranges, harming IoU-sensitive detection. The paper replaces angular regression with fine-grained classification, adds CSL and DCL, and reports strong performance across detectors and datasets.

  • Motivation: Regression-based rotation detectors suffer boundary problems because ideal predictions can lie beyond the defined range.Angle periodicity and corner ordering create discontinuities that make boundary conditions difficult.
  • Proposed direction: The proposed classification paradigm discretizes angles into intervals finer than 1 degree, making conversion error negligible.It contrasts with dominant regression-based methods and older coarse classification approaches.
  • Contributions: Circular Smooth Label handles angular periodicity and adjacent-angle tolerance, while Densely Coded Label reduces CSL’s encoding length and parameter cost.An angle fine-tuning mechanism further addresses theoretical errors caused by angle dispersion.
  • Contributions: OHDet extends rotation detection to object heading detection, and OHD-SJTU supports both rotation and heading detection tasks.The heading detector identifies an object’s head from orientation information.
  • Results: Experiments on HRSC2016 and DOTA-v1.0 report state-of-the-art detector performance and verify CSL and DCL across different detectors.The journal version also expands evaluation to additional challenging datasets and settings.

2 RELATED WORK

Prior work mainly adapts object detectors with rotated boxes or quadrilaterals, while classification-based orientation methods often lack the precision needed for aerial imagery and scene text. The paper positions its classification-based rotation baseline and heading dataset as responses to these limitations.

  • Rotated detection: Rotation detection research primarily adapts classical detectors using rotated bounding boxes or quadrilaterals for aerial images and scene text.These applications involve complex scenes, small objects, and multi-oriented instances.
  • Limitations: Regression-based rotated-box and quadrilateral methods often suffer boundary problems from predictions beyond the defined range.This motivates a baseline that transforms angle prediction from regression to classification.
  • Segmentation alternatives: Segmentation alternatives face labeling costs and dense-scene limitations, including suppression of overlapping boxes and difficulty assigning many small instances.A cited DOTA-v1.5 example contains up to 2,000 vehicles smaller than 10 pixels in a 450 × 600 sub-image.
  • Classification for orientation: Earlier face-orientation classifiers provide orientation information, but their coarse or specialized designs do not establish high-precision general rotation detection.The paper distinguishes its fine-grained classification goal from these earlier approaches.
  • Heading detection: OHD-SJTU adds labeled heading information and covers more object categories than the previously discussed heading-detection resources.The paper targets multi-class high-precision rotation and object heading detection.

3 PROPOSED APPROACH

The proposed detector is a multi-task RetinaNet-based pipeline with classification, rotation-detection, and object-heading branches. Its rotation branch can compare regression with CSL-based prediction, while the classification module can replace regression modules in many protocols.

  • Architecture: The RetinaNet embodiment contains classification, rotation-detection, and object-heading branches in a multi-task pipeline.The rotation branch includes both regression-based and CSL-based prediction for comparison.
  • Architecture: The CSL-based protocol is presented as more accurate for learning object orientation and scale information than the regression-based alternative.The figure overview also indicates that DCL retains the size-reduced encoding design.
  • Transferability: The proposed classification module is applicable to most regression-based protocols by replacing their regression module.This frames CSL/DCL as reusable prediction components rather than a detector-specific redesign.

3.1 Regression-based Rotation Detection Method

Regression-based rotation detection represents oriented objects with either five parameters and an angle or eight coordinates for ordered quadrilateral corners. These parameterizations impose different angular ranges and ordering requirements.

  • Five-parameter regression: Five-parameter methods represent a rotated rectangle with center coordinates, width, height, and angle.The paper discusses variants using 90° or 180° angular ranges.
  • Five-parameter regression: The 90° representation uses the acute angle to the x-axis, whereas the 180° representation determines the angle from the rectangle’s long side.The parameter ordering differs between the two definitions.
  • Eight-parameter regression: Eight-parameter methods directly regress the four quadrilateral corners after sorting them in advance.Corner ordering is a key step in quadrilateral regression.

3.2 Boundary Problem of Regression Method

Regression-based rotation detectors suffer boundary discontinuities across angle- and corner-based parameterizations because ideal predictions can exceed the defined representation range. Replacing regression with classification removes the observed boundary problem in the compared CSL- and DCL-based protocols.

  • Boundary discontinuities arise from angular periodicity in five-parameter protocols and corner ordering in eight-parameter protocols.
  • At boundaries, discontinuity can sharply increase the loss and make predictions inaccurate, especially when the ideal regression path lies outside the defined range.
  • The 90° protocol has periodicity and edge-exchangeability cases, while the 180° protocol still suffers angular-periodicity loss near boundaries.
  • Point regression also exhibits boundary discontinuity because fixed corner correspondences can differ from the ideal corner ordering.
  • Existing IoU-smooth L1 and modular rotation losses mitigate boundary loss but remain regression-based and do not address its root cause.
  • Classification-based CSL and DCL protocols show no boundary problem in the reported visual comparisons.

3.3 Vanilla Angular Classification

Vanilla angular classification discretizes the angle range into categories, with discretization error becoming small at fine granularity. However, prior coarse face-detection classification and boundary-case comparisons motivate more suitable angular classification designs.

  • Vanilla classification assigns each angle interval of width ω to a category, with ω = AR/Cθ linking angle range and category count.
  • For 1:9 aspect-ratio rectangles, 0.25° and 0.5° angle errors reduce IoU by only 0.02 and 0.05, respectively.
  • Earlier face-detection methods used approximate angular categories around 10°, whereas rotation detection may require finer precision for large-aspect-ratio objects and multiple categories.
  • The compared boundary-case protocols include regression-based methods and angle-classification methods, with classification protocols evaluated alongside horizontal and rotating anchors.

3.4 Circular Smooth Label for Angular Classification

Circular Smooth Label converts angle prediction into periodic classification with locally smooth labels, addressing angle-distance insensitivity and periodicity. The method uses window functions, while the 180° definition remains necessary for the edge-exchangeability problem.

  • The 90° protocol retains edge-exchangeability cases and two border types, whereas the 180° protocol has only vertical border cases.
  • Vanilla classification loss treats predictions at different angular distances from the ground truth equally, despite nearby angles being more acceptable for detection.
  • CSL provides circular, periodic labels with smooth tolerance around the target angle, improving classification-based angular prediction under boundary conditions.
  • CSL labels use a window function centered on the current bounding-box angle, with radius r determining the nonzero neighborhood.
  • An ideal CSL window is periodic, symmetric, bounded by a maximum of 1, and monotonically decreases away from its center.
  • Pulse, rectangular, triangle, and Gaussian windows satisfy the stated CSL properties; one-hot labeling is equivalent to a pulse-window or very small-radius CSL.

3.5 Densely Coded Label for Angular Classification

CSL and one-hot encodings can make prediction layers inefficient because their code lengths scale with the angular range. Densely Coded Labels reduce encoding length substantially while preserving fine angular discretization and improving detector efficiency.

  • One-hot and CSL are sparse encodings whose long code lengths increase prediction-layer parameters and computation.
  • DCL methods such as Binary Coded Label and Gray Coded Label represent more values with shorter codes than sparse encodings.
  • Table 1 compares three-digit binary and Gray codes, while Figure 7 illustrates encoding and decoding for one-hot, CSL-Gaussian, and BCL angle prediction.
  • The prediction-layer calculation depends on anchor count and shared convolutional-layer dimensions, including A = scale num × ratio num × angle num.
  • With RetinaNet settings, CSL and one-hot require code length 180, whereas DCL requires only 8, making code shortening important for prediction-layer cost.
  • RetinaNet-DCL is reported to be about 3 times and 2 times faster than RetinaNet-CSL under the compared settings.
  • For 180° divided into 256 categories, DCL uses ω = 0.703125° with maximum and expected errors of 0.3515625° and 0.17578125°.

3.6 Angle Fine-Tuning

Angle fine-tuning compensates for discretization error introduced by coarse angle classification. The method predicts a small correction ε from an additional output and applies it to the encoded angle while avoiding PoA.

  • Angle discretization: A larger angle granularity ω reduces prediction-layer parameters but introduces theoretical angle error that cannot be ignored.The method addresses this trade-off by predicting a smaller corrective angle ε.
  • Correction branch: An extra one-dimensional output εlogits is added to the prediction layer to fine-tune the predicted angle.The fine-tuning branch is illustrated using RetinaNet.
  • Angle constraint: The min and max operations constrain the fine-tuned angle to avoid PoA.
  • Heading label: The heading label identifies the coordinate quadrant pointed to by the object’s head orientation.The red dot marks the head, the green dot marks the rotating-box center, and the yellow dotted line indicates head orientation.

3.7 Object Heading Detector

The object heading detector adds fine-grained head-direction prediction to rotation detection. It uses a four-category head prediction strategy within a cascade design that combines regression and angle classification across stages.

  • Object heading detection: Object heading detection determines an object’s head, whereas rotation detection alone cannot identify the accurate head from the rotating-box angle.
  • Object heading detection: The head is assumed to lie on one of the four sides of the rotating bounding box, enabling an additional four-category prediction.
  • Design requirements: The detector is designed to provide precise boxes away from boundaries and reduce sensitivity to boundary problems at boundaries.
  • Cascade design: The cascade combines regression and angle classification, omitting angle classification in the first stage to preserve an appropriate number of anchors.
  • Cascade design: Later refined stages retain the highest-confidence anchor at each feature point, making A equal to 1 and permitting angle classification.
  • Cascade design: The 90-degree angle definition reduces aspect ratio, while IoU-Smooth L1 is used at each stage for boundary and EoE issues.
  • Results and detector: Proper adjacent-angle fault tolerance improves detector performance, and OHDet4 combines head prediction with the preceding procedures.For DOTA’s AP50 evaluation, increasing w and adjusting the window radius r can relieve prediction-layer pressure.

3.8 Loss Function Design

The loss design supports a multitask detector with regression and angle-classification branches. It defines targets for box geometry, angle error, heading, and category prediction, while combining task-specific losses with configurable trade-offs.

  • Loss formulation: The multitask pipeline contains regression-based and angle-classification-based prediction branches for direct comparison.The model outputs location and size information as part of the multitask formulation.
  • Box targets: Box targets encode center coordinates and dimensions relative to an anchor using normalized offsets and logarithmic width and height terms.The target equations use x, y, w, and h for box center coordinates, width, height, and angle, respectively.
  • Angle branches: The regression branch compares direct angle-offset regression with indirect regression using sine and cosine angle vectors.
  • Angle branches: The sine-cosine representation is normalized to satisfy cos^2 θ + sin^2 θ = 1.
  • Loss formulation: Indirect regression is incorporated into a multitask loss alongside classification, regression, angle-error, and heading terms.The formulation distinguishes theoretical angle error εgt from predicted angle error εpred and ground-truth from predicted heading.
  • Loss formulation: The loss weights λk are set by default to {1, 0.5, 20, 0.1, 1}, while classification terms use focal or sigmoid cross-entropy loss and regression uses smooth L1 loss.

4 EXPERIMENTS

Experiments evaluate classification-based angle prediction across aerial, face, and scene-text datasets, showing benefits over regression protocols and sensitivity to discretization, window design, and heading metrics.

  • 4.2 Ablation Study: Classification-based CSL, BCL, and GCL outperform the compared regression methods on DOTA-v1.0 validation experiments.The reported comparison includes direct and indirect regression protocols alongside classification-based alternatives.
  • 4.2 Ablation Study: 35.04%, 36.71% and 36.34% are reported for the compared CSL window-function results on the DOTA-v1.0 validation set, respectively.The Gaussian window performs best overall, while the pulse function performs worst; 180°-CSL improves boundary prediction over 90°-CSL.
  • 4.2 Ablation Study: A discrete granularity of 10 gives CSL its highest DOTA-v1.0 performance, while too few or too many categories reduce performance.Smaller granularity improves theoretical angular accuracy, but excessive category counts challenge classification; DCL peaks between 32 and 128 categories.
  • 4.2 Ablation Study: PCA visualizations show Gaussian CSL features forming overlapping rings for adjacent angles, whereas pulse-window features are relatively random.The reported feature structure is associated with learning orientation and scale information and eliminating boundary problems.
  • 4.3 Comparisons on Other Datasets: 89.62% (96.10%), 89.56% (96.02%), and 89.46% (96.41%) are reported for CSL, GCL, and BCL on HRSC2016.HRSC2016 contains many large-aspect-ratio ships with arbitrary orientations, making localization accuracy challenging.
  • 4.4 Object Heading Detection Experiment: OHDet reaches about 63.94% on OHD-SJTU-S, with gains of 3.93% in AP75 and 2.08% in AP50:95 over the second-best method.Object heading evaluation reports OBB mAP, OHD mAP, and Head Accuracy; OHD mAP additionally accounts for head prediction.

5 CONCLUSION

The paper identifies out-of-range ideal predictions as the root cause of boundary problems in regression-based rotation detection. It replaces angle regression with classification, adds CSL and DCL, and extends detection to object headings with the OHD-SJTU dataset and OHDet detector.

  • Ideal predictions falling outside the defined range are identified as the main cause of boundary problems in regression-based rotation detection.
  • Classifying object angles better limits predictions and allows CSL to handle angular periodicity while increasing tolerance between adjacent angles.
  • DCL greatly shortens CSL's encoding while maintaining angle prediction accuracy, and angle fine-tuning addresses dispersion errors shared by both methods.
  • OHDet detects object heads, while the released OHD-SJTU dataset supports rotation and object heading detection experiments.
Loading 2003.05597v4…