PaperlayerView exact revision on arXiv

Source-linked AI summary

Oriented Objects as pairs of Middle Lines

Haoran Wei, Yue Zhang, Zhonghan Chang, Hao Li, Hongqi Wang, Xian Sun

arXiv:1912.10694v3cs.CV

TL;DR

Oriented-object detection is challenged by redundant horizontal boxes and the computational costs of anchor- and NMS-based detectors. O2-DNet addresses this with one-stage, anchor-free, NMS-free prediction of paired middle lines, achieving competitive results across oriented, text, and natural-object benchmarks.

  • Problem

    Horizontal boxes can contain redundant background, while rotated anchors and NMS increase complexity in oriented-object detection.

  • Method

    O2-DNet detects each object through paired middle lines, combining keypoint detection, dense sampling, drift regions, two branches, and Line Loss.

  • Results

    O2-DNet is competitive with state-of-the-art detectors across aerial oriented-object, natural-scene text, and natural-object detection.

  • Takeaways & Limitations

    The model provides a single anchor-free, NMS-free network for detecting both oriented and horizontal objects without increasing computational complexity through its two-branch design.

Abstract

from arXiv · show

The detection of oriented objects is frequently appeared in the field of natural scene text detection as well as object detection in aerial images. Traditional detectors for oriented objects are common to rotate anchors on the basis of the RCNN frameworks, which will multiple the number of anchors with a variety of angles, coupled with rotating NMS algorithm, the computational complexities of these models are greatly increased. In this paper, we propose a novel model named Oriented Objects Detection Network O^2-DNet to detect oriented objects by predicting a pair of middle lines inside each target. O^2-DNet is an one-stage, anchor-free and NMS-free model. The target line segments of our model are defined as two corresponding middle lines of original rotating bounding box annotations which can be transformed directly instead of additional manual tagging. Experiments show that our O^2-DNet achieves excellent performance on ICDAR 2015 and DOTA datasets. It is noteworthy that the objects in COCO can be regard as a special form of oriented objects with an angle of 90 degrees. O^2-DNet can still achieve competitive results in these general natural object detection datasets.

1 Introduction

Oriented objects challenge horizontal-box detectors because background redundancy and dense, elongated layouts can cause missed detections. Existing approaches commonly rely on RCNN frameworks, rotated anchors, and specialized NMS, increasing computational complexity.

  • Horizontal boxes can include redundant background pixels when objects appear at substantial angles.
  • Dense, elongated objects can overlap heavily under horizontal boxes, causing NMS to miss detections.
  • Oriented bounding boxes address the redundancy and overlap problems associated with horizontal-box detection.
  • R2CNN extends Faster R-CNN with regression for oriented-box geometry, but horizontal anchors remain subject to RPN NMS suppression.
  • RRPN replaces horizontal anchors with rotation anchors and uses a corresponding rotated NMS algorithm.
  • SCRDet adds IOU loss and multidimensional attention for oriented aerial detection while remaining anchor-based and NMS-based.

middle lines

O2-DNet detects oriented objects without anchors or NMS by combining keypoint localization, dense sampling, and regression of two internal middle lines. Its branches and Line Loss address drift, angle boundaries, and line geometry.

  • middle lines: O2-DNet is a one-stage, anchor-free, NMS-free detector that represents each oriented object with a pair of middle lines.
  • middle lines: The model combines keypoint detection with dense sampling and uses a drift region so intersection-point errors within that region do not move the final box.
  • middle lines: O2-DNet uses two branches to predict 90-degree horizontal objects separately from objects at other angles.
  • middle lines: Line Loss regresses middle-line endpoints through position, parallel, and vertical components that encode their geometric relationships.

2 Related Works

Related work spans traditional feature-engineered detectors, deep convolutional detectors, and anchor-based or anchor-free formulations. Existing anchor-free methods use keypoints or point regression but can struggle with many targets in one image.

  • Traditional detectors manually engineer features such as HOG before classification with models such as SVM.
  • Deep convolutional detectors automatically extract features through backbone networks and have improved detection accuracy and robustness.
  • Anchor mechanisms originated in Faster R-CNN’s RPN and became widely used in two-stage and later one-stage detectors.
  • Anchor-free detectors locate objects with keypoints or regressions from many points, but oriented-object methods can require retaining K top-scoring objects.
  • In images containing numerous targets, retaining only K highest-scoring detections may cause missed detections, motivating O2-DNet’s combined formulation.

3 O2-DNet

O2-DNet detects objects by predicting two corresponding middle lines and their intersection point, using separate branches for horizontal and oriented objects. It regresses line endpoints relative to the intersection and adds specialized losses and a drift region to improve bounding-box extraction.

  • Middle Lines and Their Intersection Point: Each object is represented by two middle lines, with each line defined by two corresponding endpoints and an intersection point.The intersection point is predicted as a heatmap keypoint, while line endpoints are obtained through regression.
  • Architecture: O2-DNet uses two branches to predict heatmaps and regression maps for horizontal and oriented objects.The branches independently handle objects with 90-degree angles and other orientations.
  • Middle Lines and Their Intersection Point: The model regresses each endpoint’s relative distance from the intersection point to recover both middle lines.For each middle line, four regression maps predict the x and y offsets of its two endpoints.
  • Middle Lines and Their Intersection Point: A three-part Line Loss constrains endpoint positions and relationships so predicted middle lines remain geometrically consistent.The method addresses non-collinearity between endpoints and the intersection point and controls the relationship between the two lines.
  • Drift Region: The drift region lets nearby pixels regress endpoint distances, reducing dependence on exact intersection-point localization.Its radius depends on the output stride and the lengths of the two middle lines, with r fixed at 16.

4 Experiments

The experiments evaluate O2-DNet across aerial-image detection, natural-scene text detection, and general natural-object detection datasets. DOTA uses oriented bounding-box annotations and cropped aerial images, while ICDAR 2015 provides natural-scene text images.

  • Dataset selection: Three datasets cover aerial oriented-object detection, natural-scene text detection, and natural-object detection.The selected benchmarks represent different detection settings and research fields.
  • DOTA: DOTA contains 2,806 aerial images spanning 15 categories, including aircraft, small car, and ship.Image sizes range from 800 × 800 to 4000 × 4000 pixels.
  • DOTA: The experiments use DOTA’s oriented-bounding-box task rather than its horizontal-bounding-box task.Large DOTA images are cropped into 800 × 800 patches with 0.25 overlap.
  • ICDAR 2015: ICDAR 2015 contains 1,000 training images and 500 test images for natural-scene text detection.The images have a size of 720 × 1280 pixels.

COCO

COCO is used to test O2-DNet’s generality on natural-object detection with horizontal bounding-box annotations. Testing includes a specified heatmap threshold and branch-selection rule.

  • Dataset: COCO contains 80k training images, 40k validation images, and 20k testing images across 80 categories.Its annotations are horizontal bounding boxes used to test model generality.
  • Implementation: During training, the DOTA configuration uses a 511 × 511 input, output stride 4, Adam optimization, and 300k iterations.Training starts at learning rate 0.001, reduced tenfold every third interval, with batch size 32.
  • Inference: At test time, the heatmap threshold is 0.3, and the higher intersection-point score selects between outputs from both branches.Both branches may produce outputs for objects at critical angles.

4.3 Comparisons with State-of-the-art Frameworks

O2-DNet is evaluated against state-of-the-art methods on DOTA, ICDAR 2015, and COCO. It achieves strong reported results across aerial, text, and general natural-object detection.

  • DOTA: The DOTA comparison table covers 15 object categories and includes modified SSD, YOLOv2, and RetinaNet models that output oriented boxes.The listed categories include plane, bridge, ship, small vehicle, and helicopter.
  • DOTA: 71.04% mAP on DOTA is reported, exceeding most compared two-stage and one-stage aerial-object detectors.The model achieves the highest AP for bridges with large aspect ratios and densely parked small vehicles.
  • ICDAR 2015: 82.97% F1 on ICDAR 2015 is reported, exceeding the compared methods.The result supports use on natural-scene text annotations represented by irregular quadrilaterals.
  • COCO: 41.3% AP on COCO is reported, leading most one-stage detectors in the comparison.COCO objects use horizontal bounding boxes, so O2-DNet outputs only through its first branch.

4.4 Ablation Studies

The ablation study examines backbone choice, Line Loss, and branch count on DOTA. Its comparison table defines the tested backbone, loss, and single-branch variants.

  • Ablation design: Three DOTA ablations test backbone choice, Line Loss, and use of a single branch.Table 4 reports the experimental data for these three factors.
  • Backbone: The ResNet101-FPN variant replaces the model’s backbone with ResNet101-FPN.This isolates the effect of backbone selection in the ablation table.
  • Line Loss: The no-Line-Loss variant retains only the L1 component of the total Line Loss.This compares the complete line objective with a reduced loss configuration.
  • Branches: The single-branch variant keeps only O2-DNet’s second branch.This tests the contribution of the two-branch design.

ResNet101-FPN

With ResNet101-FPN, O2-DNet achieves 68.93% mAP, while Line Loss improves performance by 1.92% mAP and controls predicted median-line properties.

  • ResNet101-FPN: 68.93% mAP is achieved by O2-DNet with ResNet101-FPN.This result indicates that performance does not depend entirely on 104-Hourglass.
  • Line Loss: 1.92% mAP improvement is obtained with Line Loss compared with the model without Line Loss.The authors attribute the gain to better control of the regression target median lines’ line-segment properties.
  • Line Loss: Figure 7 compares model outputs with Line Loss against outputs without Line Loss.
  • Two branches: 2.23% higher mAP is achieved by the two-branch model than by the single-branch model.The experiment evaluates whether the two branches better address the boundary problem.

Single branch

Figure 8 shows qualitative results output by O2-DNet.

  • Single branch: Figure 8 presents qualitative results from O2-DNet.
  • Single branch: The figure’s outputs are attributed to O2-DNet.
  • Single branch: Figure 8 is a qualitative-results visualization rather than a reported quantitative evaluation.

5 Conclusion

O2-DNet detects oriented objects by predicting a pair of middle lines inside each target. The model is reported as competitive with state-of-the-art detectors across aerial imagery, natural-scene text, and natural-image object detection.

  • 5 Conclusion: O2-DNet is a one-stage, anchor-free model that detects oriented objects by predicting a pair of middle lines inside each target.
  • 5 Conclusion: The model locates each object through a pair of middle lines rather than the conventional oriented bounding-box representation.
  • 5 Conclusion: O2-DNet is competitive with state-of-the-art detectors across aerial-image oriented-object detection, natural-scene text detection, and natural-image object detection.
Loading 1912.10694v3…