PaperlayerView exact revision on arXiv

Source-linked AI summary

Fourier-based Rotation-invariant Feature Boosting: An Efficient Framework for Geospatial Object Detection

Xin Wu, Danfeng Hong, Jocelyn Chanussot, Yang Xu, Ran Tao, Yue Wang

arXiv:1905.11074v1cs.CV

TL;DR

Geospatial object detection must handle object deformations, including rotation, that challenge existing detectors. FRIFB combines FourierHOG, aggregate channel features, FPGM, and boosting; experiments on two NWPU VHR-10 subsets report competitive performance against prior methods.

  • Problem

    Scaling, offset, and rotation deform remote-sensing objects, and existing approaches have difficulty capturing rotation-related properties, particularly with small training samples.

    Source: §I · p. 1

  • Method

    FRIFB integrates rotation-invariant FourierHOG features, aggregate channel features, FPGM pyramid generation, and boosting.

    Source: B. Contributions · p. 2

  • Results

    FRIFB shows competitive performance against previous state-of-the-art algorithms on two NWPU VHR-10 subsets.

    Source: B. Contributions · p. 2

  • Takeaways & Limitations

    The authors report that FRIFB is robust to rotation and can improve detection performance in the evaluated geospatial object-detection setting.

    Source: §IV · p. 5

Abstract

from arXiv · show

Geospatial object detection of remote sensing imagery has been attracting an increasing interest in recent years, due to the rapid development in spaceborne imaging. Most of previously proposed object detectors are very sensitive to object deformations, such as scaling and rotation. To this end, we propose a novel and efficient framework for geospatial object detection in this letter, called Fourier-based rotation-invariant feature boosting (FRIFB). A Fourier-based rotation-invariant feature is first generated in polar coordinate. Then, the extracted features can be further structurally refined using aggregate channel features. This leads to a faster feature computation and more robust feature representation, which is good fitting for the coming boosting learning. Finally, in the test phase, we achieve a fast pyramid feature extraction by estimating a scale factor instead of directly collecting all features from image pyramid. Extensive experiments are conducted on two subsets of NWPU VHR-10 dataset, demonstrating the superiority and effectiveness of the FRIFB compared to previous state-of-the-art methods.

I. INTRODUCTION

Scaling, offset, and rotation in remote-sensing imagery degrade geospatial object detection. The paper motivates continuous rotation-invariant features and efficient pyramid extraction to address limitations in prior approaches.

I. INTRODUCTION

  • Scaling, offset, and rotation deform remote-sensing objects and degrade geospatial object detection performance.Object rotation is also identified as an important factor in degraded detection.Source: §I · p. 1, A. Motivation · p. 1
  • Prior methods often fail to capture rotation-related properties with small training samples, while manually designed descriptors can be costly and limited by discrete coordinates.Learned features depend on the quantity and diversity of augmented samples; dense manual extraction can be expensive.Source: §I · p. 1
  • The rotation challenge comprises rotation-invariant feature extraction and image-pyramid generation, with repeated feature extraction across pyramid layers adding computational cost.FourierHOG addresses the feature-coordinate limitation, while FPGM estimates a scale factor for parallel pyramid feature extraction.Source: §I · p. 1

A. Motivation

  • FPGM is fast, but its low-level shift-invariant input requirement motivates a more discriminative, rotation-robust descriptor.The paper identifies FourierHOG as a candidate that can be embedded in FPGM.Source: A. Motivation · p. 2

B. Contributions

FRIFB combines FourierHOG, aggregate channel features, FPGM, and boosting for geospatial object detection. The paper reports competitive performance on two NWPU VHR-10 subsets.

B. Contributions

  • FRIFB combines rotation-invariant FourierHOG, aggregate channel features, FPGM, and boosting with decision-tree ensembles.The framework assigns feature extraction, feature refinement, pyramid generation, and learning to these components.Source: B. Contributions · p. 2
  • FourierHOG's rotation and shift robustness fits FPGM's assumptions, while FPGM provides faster pyramid-feature computation.The authors describe the components' complementary advantages as supporting fast and robust detection.Source: B. Contributions · p. 2
  • FRIFB is evaluated on baseball and airplane subsets of NWPU VHR-10 and is reported to show competitive performance against prior state-of-the-art algorithms.The work describes both qualitative and quantitative evaluation.Source: B. Contributions · p. 2

II. METHODOLOGY

FRIFB trains on fixed-size windows, generates FourierHOG features, refines them with ACF, and trains decision trees through boosting. At test time, it uses FPGM for pyramid-feature generation.

II. METHODOLOGY

  • Training begins with a fixed bounding-box size, with targets upsampled or downsampled to that size across scales.The size is generally assigned by averaging the training samples.Source: A. Overview · p. 2
  • FourierHOG generates rotation-invariant channel maps from the training windows.The procedure applies gradient magnitude and phase processing over Fourier orders before region convolution.Source: A. Overview · pp. 2–3
  • The training procedure learns decision trees over multiple boosting rounds, and testing uses FPGM to generate pyramid features.The workflow specifies octave-paced scale intervals at test time.Source: A. Overview · pp. 2–3
  • ACF structurally refines the extracted rotation-invariant channel features before boosting training.The overview identifies ACF features as input to the boosting stage.Source: A. Overview · p. 3

B. Rotation-invariant Feature Generation (FourierHOG)

The section constructs rotation-invariant features by representing gradient information in polar-coordinate Fourier form and combining features so their rotation behavior cancels. The resulting convolutional features are presented as the final rotation-invariant representation.

B. Rotation-invariant Feature Generation (FourierHOG)

  • Polar coordinates separate angular and radial information, with Fourier basis functions representing angular information by rotation order.The basis is ψ_m(ϕ) = e^imϕ.Source: B. Rotation-invariant Feature… · p. 3
  • Gradient magnitude and direction are encoded as a continuous angular impulse, whose Fourier representation supplies the feature components.The impulse uses the gradient magnitude and phase angle at each pixel.Source: B. Rotation-invariant Feature… · p. 3
  • Combining Fourier features whose rotation orders sum to zero yields convolutional features identified as the final rotation-invariant representation.The construction uses self-steerability kernels obtained through inverse Fourier transformation.Source: B. Rotation-invariant Feature… · p. 3

C. Aggregate Channel Features (ACF)

Aggregate channel features (ACF) refine rotation-invariant channel maps through subsampling, creating representations at different levels of structure.

C. Aggregate Channel Features (ACF)

D. Fast Pyramid Generative Model (FPGM)

FPGM estimates feature scaling across image scales to construct finely sampled pyramids with octave-spaced intervals. The paper embeds Fourier-based rotation-invariant feature maps in this framework.

D. Fast Pyramid Generative Model (FPGM)

A. Data Description and Experiment Setup

The experiments evaluate FRIFB on airplane and baseball-diamond scenes from NWPU VHR-10, using a training/test split and five-fold cross-validation.

A. Data Description and Experiment Setup

  • The benchmark uses airplane and baseball-diamond scenes from NWPU VHR-10, with 60% of scene images for training and the remainder for testing.Training positives are mirror-augmented; negatives are randomly selected from 150 target-free images.Source: A. Data Description and Experiment Setup · p. 4
  • Experiments use five-fold cross-validation and report average results, with average target sizes of 75 × 75 for airplanes and 90 × 90 for baseball diamonds.Source: A. Data Description and Experiment Setup · p. 4
  • Performance is evaluated with precision-recall curves and average precision, counting detections as true positives when box IoU exceeds 0.5.The benchmark imagery includes Google Earth data at 0.5–2 m spatial resolution and infrared imagery at 0.08 m.Source: A. Data Description and Experiment Setup · p. 4

B. Detection on NWPU VHR-10 dataset

On NWPU VHR-10, FRIFB’s visual detections are described as robust to rotation, and comparisons report stronger precision alongside a speed–accuracy trade-off.

B. Detection on NWPU VHR-10 dataset

  • FRIFB detects most airplane and baseball-diamond targets with fewer false positives, though discrimination remains limited around airplane tail fins and field edges or corners.The authors say robustness depends on training-sample resolution and Fourier basis-function settings.Source: B. Detection on NWPU VHR-10 dataset · p. 4
  • Robustness to rotations and complex noise is theoretically guaranteed when training samples and Fourier basis functions are sufficiently sampled.This condition qualifies the robustness claim for the detections discussed in the section.Source: B. Detection on NWPU VHR-10 dataset · p. 4

C. Comparison with State-of-the-Art Algorithms

  • The comparison includes BOW-SVM, rotation-aware features, Exemplar-SVMs, COPD, R-DFPN, RICNN, YOLO2, and FPGM, with rotated or translated training augmentation across methods.The authors refer readers to the cited works for parameter settings.Source: C. Comparison with State-of-the-Art… · p. 4
  • Around 1% and 5%: YOLO2’s detection performance is below FRIFB on the two datasets; FRIFB also outperforms other methods in precision.YOLO2 is described as the fastest, while the authors suggest FRIFB may suit applications prioritizing accuracy with acceptable running time.Source: C. Comparison with State-of-the-Art… · p. 4

IV. CONCLUSION

The conclusion presents FRIFB as a rotation-invariant remedy for fast-pyramid sensitivity and reports robust detection with improved performance. It also identifies tiny-object detection as future work.

IV. CONCLUSION

  • FRIFB addresses the rotation sensitivity of fast pyramid features by introducing a rotation-invariant descriptor.The descriptor is integrated into a framework combining techniques with complementary advantages.Source: §IV · pp. 4–5
  • Table I compares methods using AP values and average running time per image.Its caption identifies the evaluation measures but does not supply the comparison values.Source: Table I · p. 5
  • Fig. 3 shows proposed-method detection results for airplanes and baseball diamonds in NWPU VHR-10.Source: Figure 3 · p. 5
  • The authors report that FRIFB is robust to rotation and effectively improves detection performance.The conclusion does not provide quantitative results in the supplied passage.Source: §IV · p. 5
  • Future work will target tiny-object detection through end-to-end learning or auxiliary hyperspectral or multispectral data.Source: §IV · p. 5

Related papers

Loading 1905.11074v1…