PaperlayerView exact revision on arXiv

Source-linked AI summary

Dynamic Anchor Learning for Arbitrary-Oriented Object Detection

Qi Ming, Zhiqiang Zhou, Lingjuan Miao, Hongwei Zhang, Linhao Li

arXiv:2012.04150v2cs.CV

TL;DR

IoU-based anchor assignment can select poorly localizing positives and overlook accurate negatives, creating inconsistency between classification confidence and localization quality. The paper proposes Dynamic Anchor Learning, which uses matching degree, dynamic selection, and matching-sensitive loss to address this mismatch. Across several arbitrary-oriented and horizontal-box detection datasets, the method reports stable, substantial improvements and no added inference cost, while using a simple regression-uncertainty penalty form.

  • Problem

    IoU-based assignment does not reliably assess anchor localization quality, leaving some accurate negative candidates unused and creating classification–localization inconsistency.

  • Method

    Dynamic Anchor Learning uses matching degree, dynamic anchor selection, and matching-sensitive loss to evaluate localization potential and improve label assignment and classification–regression alignment.

  • Results

    Stable, substantial improvements are reported on HRSC2016, DOTA, UCAS-AOD, and ICDAR 2015, with applicability to horizontal-box detection and no added inference cost.

  • Takeaways & Limitations

    DAL is presented as a general detection-pipeline strategy that can improve high-precision arbitrary-oriented detection and integrate into existing systems.

  • Takeaways & Limitations

    The regression-uncertainty penalty uses a simple naive form because detection performance is reported as insensitive to the form of u.

Abstract

from arXiv · show

Arbitrary-oriented objects widely appear in natural scenes, aerial photographs, remote sensing images, etc., thus arbitrary-oriented object detection has received considerable attention. Many current rotation detectors use plenty of anchors with different orientations to achieve spatial alignment with ground truth boxes, then Intersection-over-Union (IoU) is applied to sample the positive and negative candidates for training. However, we observe that the selected positive anchors cannot always ensure accurate detections after regression, while some negative samples can achieve accurate localization. It indicates that the quality assessment of anchors through IoU is not appropriate, and this further lead to inconsistency between classification confidence and localization accuracy. In this paper, we propose a dynamic anchor learning (DAL) method, which utilizes the newly defined matching degree to comprehensively evaluate the localization potential of the anchors and carry out a more efficient label assignment process. In this way, the detector can dynamically select high-quality anchors to achieve accurate object detection, and the divergence between classification and regression will be alleviated. With the newly introduced DAL, we achieve superior detection performance for arbitrary-oriented objects with only a few horizontal preset anchors. Experimental results on three remote sensing datasets HRSC2016, DOTA, UCAS-AOD as well as a scene text dataset ICDAR 2015 show that our method achieves substantial improvement compared with the baseline model. Besides, our approach is also universal for object detection using horizontal bound box. The code and models are available at https://github.com/ming71/DAL.

Introduction

The paper identifies a mismatch between IoU-based anchor assignment and localization quality, then proposes Dynamic Anchor Learning to select anchors by localization potential and reduce classification–regression inconsistency.

  • Introduction: IoU-based assignment can misjudge anchor quality: some high-input-IoU anchors regress poorly, while poorly aligned or negative anchors may localize accurately.This challenges the assumption that spatial alignment before regression reliably predicts detection quality.
  • Introduction: Input-IoU assignment introduces classification–localization inconsistency because classification confidence follows input IoU more closely than final localization quality.High-output-IoU regression boxes can be misjudged as background, weakening confidence-based selection during inference.
  • Introduction: Dynamic Anchor Learning defines matching degree to assess anchor localization potential using spatial alignment, localization ability, and regression uncertainty.The method uses this metric for training sample selection rather than relying only on anchor–ground-truth IoU.
  • Introduction: Dynamic anchor selection removes false-positive samples and mines potential high-quality candidates, while matching-sensitive loss improves classifier discrimination for well-localized proposals.Together, these components target both label assignment and the classification–regression gap.
  • Introduction: Experiments on remote sensing and scene-text datasets show stable, substantial gains for arbitrary-oriented detection and applicability to horizontal-box detection.The approach can be integrated into existing detection pipelines without increasing inference computational cost.

Related Work

Prior detectors use horizontal or rotated anchors and sampling strategies, but their anchor-quality judgments and IoU-based assignments remain imperfect, motivating more reliable assessment.

  • Arbitrary-Oriented Object Detection: Existing rotation detectors commonly extend horizontal-box detectors with rotated anchors and additional angle prediction.These approaches retain label-assignment strategies used in general object detection frameworks.
  • Arbitrary-Oriented Object Detection: Although existing approaches perform well, they may misjudge anchor quality, causing improper label assignment that adversely affects training.The stated issue concerns the reliability of anchor assessment rather than the absence of detection performance.
  • Arbitrary-Oriented Object Detection: Dense anchor configurations create severe class imbalance, especially for arbitrary-oriented objects with an additional angle dimension.Sampling strategies and Focal Loss are described as common responses to this imbalance.
  • Arbitrary-Oriented Object Detection: Dynamic R-CNN and ATSS adjust IoU thresholds, but they do not assess whether IoU itself is a credible label-assignment criterion.Other work models label noise or cleanliness, yet the reviewed methods do not fully address the reliability problem identified here.

Proposed Method

The proposed method replaces IoU-only anchor assessment with matching degree, which combines spatial alignment, feature alignment, and regression uncertainty for dynamic sample selection. Matching-sensitive losses then weight positives by localization potential to improve classification–localization consistency.

  • Rotation Detector Built on RetinaNet: RetinaNet uses predefined horizontal anchors across feature levels P3–P7 and regresses center, size, and angle offsets for oriented objects.The baseline is a one-stage detector with ResNet-50 and an FPN-like multiscale feature pyramid.
  • Dynamic Anchor Selection: IoU-based assignment is unreliable because 26% of positive anchors are poorly aligned after regression, while some unmatched negatives have high localization potential.Input IoU does not fully determine localization performance, and regression uncertainty can mislead assignment.
  • Dynamic Anchor Selection: Matching degree combines input IoU, regression-box IoU, and a regression-uncertainty penalty to measure an anchor’s localization capacity.It is defined as md = α · sa + (1 −α) · fa −u^γ, where sa is spatial alignment, fa is feature alignment, and u measures IoU variation before and after regression.
  • Dynamic Anchor Selection: Anchors exceeding a matching-degree threshold of 0.6 become positives, while the highest-scoring anchor is added when a ground truth matches none.This dynamic selection is intended to remove false positives and mine high-quality candidates while reducing regression-uncertainty interference.
  • Matching-Sensitive Loss: The matching-sensitive loss weights positive classification and regression contributions by a compensation factor based on matching degree.The weighting distinguishes positives with different localization potential so high-quality samples receive greater classifier attention and regression influence.
  • Matching-Sensitive Loss: MSL is designed to strengthen the correlation between classification score and regression-box localization quality, addressing unreliable confidence-based prediction selection.Figure 3 compares this correlation with and without MSL using output IoU and classification score.

Datasets

Experiments use oriented bounding-box datasets spanning remote sensing and scene text detection. The datasets include HRSC2016, DOTA, and UCAS-AOD, with DOTA providing 15 categories and 188,282 annotated instances.

  • HRSC2016 is a challenging remote sensing ship-detection dataset containing 1061 images.
  • DOTA contains 2806 aerial images, 188,282 annotated instances, and 15 object categories.
  • DOTA images are cropped into 800×800 patches with a stride of 200 because the original images are too large.
  • UCAS-AOD is an aerial aircraft and car detection dataset containing 1510 images.

Implementation Details

The baseline uses RetinaNet with only a few horizontal anchors and standardized 800×800 inputs. Training uses augmentation, Adam optimization, scheduled learning-rate decay, and dataset-specific iteration budgets.

  • The baseline sets three horizontal anchors with aspect ratios {1/2, 1, 2} for HRSC2016, DOTA, and UCAS-AOD.
  • ICDAR uses five horizontal anchors with aspect ratios {1/5, 1/2, 1, 2, 5}.
  • All images are resized to 800×800, with random flip, rotation, and HSV colour-space transformations for augmentation.
  • Adam training starts at a learning rate of 1e-4, divided by 10 at each decay step, using batch size 8 on RTX 2080Ti.
  • Training runs for 20k, 30k, 15k, and 40k iterations on HRSC2016, DOTA, UCAS-AOD, and ICDAR 2015, respectively.

Ablation Study

Component-wise experiments on HRSC2016 show that output IoU alone is unreliable for label assignment, while suppressing regression uncertainty enables effective alignment and improves detection.

  • Output IoU reduces detection performance from 80.8% to 78.9% even with α set to 0.8 for stable training.
  • Suppressing regression uncertainty allows prior space alignment and posterior feature alignment to work together effectively for label assignment.
  • 4.8% higher than the baseline, the combined alignment approach substantially improves performance when regression uncertainty is suppressed.
  • Matching sensitivity loss produces an mAP of 88.6% on HRSC2016.
  • Reducing α appropriately increases the influence of feature alignment and raises mAP, whereas extremely large α causes a sharp performance decrease.

Experiment Results

DAL improves arbitrary-oriented detection across remote-sensing and scene-text benchmarks, while using few horizontal anchors and retaining applicability to horizontal-box detection.

  • DOTA: 71.44% mAP on DOTA surpasses the baseline by 3%.Integrated with DAL, vanilla RetinaNet competes with advanced methods.
  • DOTA: 76.95% mAP on DOTA improves S2A-Net performance by 2.83%, achieving the best result among compared models.
  • HRSC2016: 89.77% mAP on HRSC2016 is achieved with ResNet-101 and 800×800 inputs.With ResNet-50 and 416×416 inputs, the method reaches 88.6% mAP.
  • Efficiency: DAL uses only three horizontal anchors per position yet outperforms frameworks with many rotated anchors.With 416×416 inputs, the one-stage model reaches 34 FPS on an RTX 2080 Ti GPU.
  • UCAS-AOD: 2.3% improvement over baseline on UCAS-AOD accompanies significantly improved small-vehicle detection and stronger AP75.The authors associate the AP75 improvement with greater attention to high-quality samples.
  • ICDAR 2015: 81.5% F-measure on ICDAR 2015 improves the baseline by 4%, and multi-scale training and testing reaches 82.4%.DAL remains below some state-of-the-art text detectors because it does not specifically address long texts mistaken for multiple short texts.
  • Horizontal-box detection: Experiments on ICDAR 2013, NWPU VHR-10, and VOC2007 support DAL for horizontal-box object detection.

Conclusion

The paper proposes dynamic anchor learning for high-performance arbitrary-oriented object detection and reports its effectiveness and universality across several datasets.

  • DAL constructs matching degree from spatial alignment, feature alignment ability, and regression uncertainty for label assignment.Dynamic anchor selection and matching-sensitive loss are integrated to improve high-precision detection and reduce classification–regression divergence.
Loading 2012.04150v2…