PaperlayerView exact revision on arXiv

Source-linked AI summary

Dense Label Encoding for Boundary Discontinuity Free Rotation Detection

Xue Yang, Liping Hou, Yue Zhou, Wentao Wang, Junchi Yan

arXiv:2011.09670v4cs.CVcs.AI

TL;DR

Rotation detection is important across aerial images, scene text, and faces, yet classification-based orientation estimation remains relatively less studied and regression methods face boundary discontinuity. The paper introduces Densely Coded Labels and Angle Distance and Aspect Ratio Sensitive Weighting, reporting competitive accuracy and efficiency, including 77.37% vs. 76.17% on DOTA.

  • Problem

    Regression-based rotation detectors encounter boundary discontinuity, while classification-based orientation estimation remains relatively less studied despite rotation detection’s utility across visual applications.

  • Method

    The paper replaces Sparsely Coded Labels with Binary Coded Labels and Gray Coded Labels, then adds Angle Distance and Aspect Ratio Sensitive Weighting for angular-distance and aspect-ratio sensitivity.

  • Results

    The approach shows competitive performance regarding accuracy and efficiency across detectors and datasets, including 77.37% vs. 76.17% on DOTA dataset.

  • Takeaways & Limitations

    DCL-based classification detectors provide a lighter prediction layer with notably accelerated training, while ADARSW further improves performance for objects with small aspect ratios.

  • Takeaways & Limitations

    GCL and BCL are agnostic or partially agnostic to angle distance, because encoding differences for angles with large differences are not necessarily significant.

Abstract

from arXiv · show

Rotation detection serves as a fundamental building block in many visual applications involving aerial image, scene text, and face etc. Differing from the dominant regression-based approaches for orientation estimation, this paper explores a relatively less-studied methodology based on classification. The hope is to inherently dismiss the boundary discontinuity issue as encountered by the regression-based detectors. We propose new techniques to push its frontier in two aspects: i) new encoding mechanism: the design of two Densely Coded Labels (DCL) for angle classification, to replace the Sparsely Coded Label (SCL) in existing classification-based detectors, leading to three times training speed increase as empirically observed across benchmarks, further with notable improvement in detection accuracy; ii) loss re-weighting: we propose Angle Distance and Aspect Ratio Sensitive Weighting (ADARSW), which improves the detection accuracy especially for square-like objects, by making DCL-based detectors sensitive to angular distance and object's aspect ratio. Extensive experiments and visual analysis on large-scale public datasets for aerial images i.e. DOTA, UCAS-AOD, HRSC2016, as well as scene text dataset ICDAR2015 and MLT, show the effectiveness of our approach. The source code is available at https://github.com/Thinklab-SJTU/DCL_RetinaNet_Tensorflow and is also integrated in our open source rotation detection benchmark: https://github.com/yangxue0827/RotationDetection.

1. Introduction

The paper targets boundary discontinuity and efficiency limitations in rotation detection by replacing angle regression and sparse classification labels with DCL and ADARSW.

  • Motivation: Regression-based angle prediction suffers boundary discontinuity from angle periodicity and bounding-box edge exchangeability.These issues arise because ideal predictions can fall outside the defined range, sharply increasing boundary loss.
  • Limitations of prior classification: CSL-based classification removes boundary discontinuity but retains a heavy prediction layer and difficulty with square-like objects.The long-side definition also produces relatively large losses for nearly square boxes whose angles differ substantially despite near-identical IoU.
  • Proposed method: ADARSW makes DCL-based detectors sensitive to angular distance and object aspect ratio, improving robustness for small-aspect-ratio objects.It is designed to avoid reintroducing periodicity while adapting weighting to square-like objects.
  • Proposed method: DCL replaces SCL with Binary Coded Label and Gray Coded Label, reducing code length while improving training speed and detection accuracy.The paper reports an empirically observed training speed boost of about three times, especially for BCL.
  • Results: 77.37% vs. 76.17% on DOTA dataset is reported against a state-of-the-art CSL-based detector.The paper attributes this comparison to extensive experiments and visual analysis across different datasets and detectors.

2. Related Work

Related work adapts conventional detectors to oriented boxes for aerial imagery and scene text, while classification-based angle encoding addresses regression boundary discontinuity.

  • Horizontal Object Detection: Horizontal detectors cannot provide accurate orientation and scale information needed for some specialized visual scenarios.The related work covers two-stage, single-stage, anchor-free, and transformer-based horizontal detection paradigms.
  • Rotation Object Detection: Rotation detectors commonly represent multi-oriented objects with rotated bounding boxes or quadrangles in aerial imagery and scene text.Representative methods include ICN, ROI-Transformer, SCRDet, R3Det, CSL, RRPN, TextBoxes++, RRD, and FOTS.
  • Rotation Object Detection: Most earlier arbitrary-oriented methods regress angles and therefore overlook boundary discontinuity, while CSL transforms angle prediction into classification.CSL is presented as a boundary-discontinuity-free rotation detector, though it has efficiency and square-like-object limitations.

3. Mitigating Boundary Discontinuity by Classification

The paper characterizes boundary discontinuity as a regression failure near representation boundaries and motivates classification as a more direct remedy.

  • Boundary discontinuity: Boundary discontinuity is a sharp loss increase that prevents regression detectors from using the same prediction form at boundaries and non-boundaries.Its causes are linked to object bounding-box definitions and ideal predictions exceeding the defined range.
  • Bounding-box definitions: The five-parameter 90° representation mainly encounters angle periodicity and edge exchangeability.This representation is identified with the OpenCV definition method.
  • Bounding-box definitions: The five-parameter 180° long-side representation mainly suffers from angle periodicity.The eight-parameter representation instead involves sorting the four corners.
  • Mitigation: Boundary errors can yield inaccurate angle predictions, motivating loss constraints and alternative angle-prediction forms.The paper argues that a genuinely boundary-discontinuity-free detector would provide a more robust high-performance baseline.

4. Proposed Method

The method replaces sparsely coded angle labels with compact binary or Gray-coded labels, then adds angle- and aspect-ratio-sensitive weighting to address classification inefficiency and square-like objects.

  • Rethinking on Sparsely Coded Label Encoding: CSL-based angle classification removes boundary discontinuity but uses a heavy prediction layer and struggles with small-aspect-ratio objects.With A = 21, AR = 180, and w = 1, CSL and One-Hot require prediction-layer thickness 3,780 versus 21 for regression.
  • Rethinking on Sparsely Coded Label Encoding: Square-like boxes can have nearly 1 IoU yet produce large angle-related loss under the long-side definition.The paper attributes the discrepancy to the angle parameter and notes that the issue becomes less noticeable as aspect ratio increases.
  • Densely Coded Label: Densely Coded Labels use Binary Coded Label and Gray Coded Label encodings to represent angle categories with shorter codes than CSL and One-Hot.Under the same setting, DCL coding length is 168, while BCL and GCL prediction layers scale with the logarithm of the number of angle categories.
  • Densely Coded Label: DCL reduces computational overhead without lowering detection performance, with GFlops and parameter increases of only 3.24% and 0.92%.The reported DCL training time is almost the same as the regression-based method, contrasting with the longer CSL training time.
  • Densely Coded Label: BCL and GCL are partly insensitive to angular distance, because their code differences do not consistently reflect differences between angle values.Gray coding changes one digit between adjacent codes, but angles with large differences can still have similar codes.
  • Angle Distance and Aspect Ratio Sensitive Weighting: ADARSW combines periodic angle-distance handling with aspect-ratio sensitivity to improve robustness for small-aspect-ratio objects.The angle-distance weight uses the decimal difference between predicted and ground-truth angles, while the paper identifies long-side definitions as unsuitable for square-like objects.

5. Experiments

Experiments across aerial-image and scene-text benchmarks evaluate DCL encoding, ADARSW, angle discretization, and portability across detectors. Classification-based detectors outperform regression-based alternatives, while DCL improves efficiency and accuracy, and ADARSW strengthens performance on small-aspect-ratio objects.

  • Ablation study: Classification-based detectors gain about 1.5%-2.3% in mAP50 and 3.6%-5.5% in 5-mAP50 over RetinaNet-Reg, while DCL is nearly three times faster than CSL.DCL further improves performance by about 2% in 5-mAP50 and 0.8% in mAP50.
  • Ablation study: Smaller angle-discretization granularity ω raises the theoretical performance bound but increases category count, whereas too many categories slightly reduce performance.Figure 5 compares angle estimates under different ω values; too few categories cause a sharp theoretical accuracy loss.
  • Ablation study: DCL remains effective with redundant invalid codes: BCL reaches 36.35% when 180 categories include about 76 invalid codings, with only slight degradation after further code-length increases.DCL requires category counts that are powers of 2 for every code to map to a valid angle, but invalid codes can still be tolerated empirically.
  • Ablation study: ADARSW raises ten-category performance by 2.3% for BCL and 2.13% for GCL, increasing overall performance to 67.39% and 67.02%, respectively.The comparison targets objects with small aspect ratios under the same environment and hyperparameters.
  • More datasets and detectors: DCL outperforms CSL in most comparisons across additional datasets, improves R3Det by 1.8%, and reaches about 74.06% with RetinaNet-DCL-ResNet152 and 77.37% with R3Det-DCL-ResNet152.On HRSC2016, the model reports about 89.46% (96.41%) under the stated evaluation metrics.
  • Visual analysis: PCA visualizations with four and eight angle categories support the feasibility of classification-based orientation estimation using cross-entropy loss.Figure 6 visualizes angular features, with red dotted lines dividing categories.

6. Conclusion

The paper advances classification-based rotation detection with compact DCL encodings and ADARSW. Across datasets and detectors, these techniques provide competitive accuracy and efficiency.

  • 6. Conclusion: DCL shortens the prediction code length, making the prediction layer lighter and notably accelerating training relative to SCL-based orientation classification detectors.The paper presents two DCL techniques for the prediction layer.
  • 6. Conclusion: ADARSW makes DCL-based detectors sensitive to angular distance and object aspect ratio, further improving performance.The technique addresses sensitivity to both angular distance and object shape.
  • 6. Conclusion: Experiments on different detectors and datasets show competitive performance in both accuracy and efficiency.
Loading 2011.09670v4…