[Paper Review] Dense Label Encoding for Boundary Discontinuity Free Rotation Detection

Xue Yang, Liping Hou+3 more|arXiv (Cornell University)|Nov 19, 2020

Advanced Image and Video Retrieval TechniquesComputer Science66 references18 citations

TL;DR

This paper proposes Dense Label Encoding (DLE) for rotation detection using angle classification to eliminate boundary discontinuity issues inherent in regression-based methods. By introducing two Densely Coded Labels (DCL) and Angle Distance and Aspect Ratio Sensitive Weighting (ADARSW), the method accelerates training by ~3x and improves mAP by up to 2.3% on challenging datasets like DOTA and HRSC2016, achieving state-of-the-art performance with enhanced accuracy and efficiency.

ABSTRACT

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.

Motivation & Objective

• To address boundary discontinuity in rotation detection caused by periodic angle regression and edge exchangeability.
• To improve training efficiency and detection accuracy of classification-based rotation detectors.
• To develop a lightweight, robust angle classification framework that avoids the pitfalls of sparsely coded labels (SCL) and regression.
• To enhance sensitivity to angular distance and aspect ratio, especially for square-like and high-aspect-ratio objects.
• To establish a generalizable, high-performance solution compatible with existing detectors like RetinaNet and R3Det.

Proposed method

• Proposes two Densely Coded Labels (DCL) — Binary Code Labeling (BCL) and Gray Code Labeling (GCL) — to replace Sparsely Coded Labels (SCL), reducing code length and accelerating training.
• Introduces Angle Distance and Aspect Ratio Sensitive Weighting (ADARSW), a loss re-weighting strategy that assigns higher weights to predictions closer in angular distance and to objects with extreme aspect ratios.
• Uses the long-side definition (180° range) for bounding boxes to inherently resolve edge exchangeability (EoE) and periodicity (PoA) issues.
• Employs cross-entropy loss with DCL to enable end-to-end training, where each label corresponds to a unique angle category with dense, continuous coding.
• Applies PCA to visualize angular feature representations, confirming that DCL enables meaningful, discriminative feature learning.
• Integrates DCL into RetinaNet and R3Det, demonstrating compatibility and performance gains across multiple architectures.

Experimental results

Research questions

• RQ1 Can dense coding of angle labels significantly reduce training time compared to sparsely coded labels in classification-based rotation detection?
• RQ2 Does the proposed ADARSW loss re-weighting strategy improve detection accuracy, especially for square-like and high-aspect-ratio objects?
• RQ3 Can DCL-based detectors achieve state-of-the-art performance while remaining free of boundary discontinuity issues?
• RQ4 How does the DCL method compare to regression-based and SCL-based methods in terms of mAP and training speed across diverse datasets?
• RQ5 To what extent is the DCL framework generalizable across different detector backbones and datasets?

Key findings

• The DCL-based method achieves a 3x increase in training speed compared to SCL-based models, empirically observed across benchmarks.
• On the DOTA validation set, the DCL-based RetinaNet achieves 36.71% mAP50:95, outperforming the SCL-based CSL model (35.04%) and regression-based baselines.
• With ADARSW, the BCL-based model improves 10-category mAP by 2.3%, and the GCL-based model by 2.13% on ICDAR2015.
• On HRSC2016, the R3Det-DCL model achieves 89.46% mAP (07) and 96.41% mAP (12), setting a new state-of-the-art.
• The method is highly generalizable, improving R3Det by 1.8% mAP under large backbone and data augmentation settings.
• Visualizations via PCA confirm that DCL enables meaningful angular feature learning, even with minimal loss functions.