Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence

Xue Yang; Xiaojiang Yang; Jirui Yang; Qi Ming; Wentao Wang; Qi Tian; Junchi Yan

Source-linked AI summary

Learning High-Precision Bounding Box for Rotated Object Detection via Kullback-Leibler Divergence

Xue Yang, Xiaojiang Yang, Jirui Yang, Qi Ming, Wentao Wang, Qi Tian, Junchi Yan

arXiv:2106.01883v5Jun 3, 2021cs.CVcs.AIcs.LG

TL;DR

Existing rotated detectors inherit regression designs that limit high-precision detection, particularly for large-aspect-ratio objects. The paper introduces KLD-based regression over Gaussian bounding-box representations within a deductive framework, with horizontal detection as a special case. Gradient analysis and experiments across seven datasets and two detectors support adaptive optimization, scale invariance, and new state-of-the-art rotation-detection performance.

Problem
Existing rotated detectors often struggle with high-precision detection because current regression losses inadequately handle object-dependent parameter sensitivity, especially for large-aspect-ratio objects.
Method
The paper represents each rotated bounding box as a 2-D Gaussian and uses KLD between predicted and ground-truth distributions as a unified regression loss that degenerates to l_n-norm loss for horizontal detection.
Results
Experiments on seven public datasets with two popular detectors show effectiveness and new state-of-the-art performance for rotation detection.
Takeaways & Limitations
KLD provides object-dependent gradient adjustment and scale invariance, properties the paper identifies as desirable for robust high-precision detection across rotation, size, and aspect ratio.
Takeaways & Limitations
The method may be applied in sensitive fields including remote sensing, aviation, and unmanned aerial vehicles.

Abstract

from arXiv · show

Existing rotated object detectors are mostly inherited from the horizontal detection paradigm, as the latter has evolved into a well-developed area. However, these detectors are difficult to perform prominently in high-precision detection due to the limitation of current regression loss design, especially for objects with large aspect ratios. Taking the perspective that horizontal detection is a special case for rotated object detection, in this paper, we are motivated to change the design of rotation regression loss from induction paradigm to deduction methodology, in terms of the relation between rotation and horizontal detection. We show that one essential challenge is how to modulate the coupled parameters in the rotation regression loss, as such the estimated parameters can influence to each other during the dynamic joint optimization, in an adaptive and synergetic way. Specifically, we first convert the rotated bounding box into a 2-D Gaussian distribution, and then calculate the Kullback-Leibler Divergence (KLD) between the Gaussian distributions as the regression loss. By analyzing the gradient of each parameter, we show that KLD (and its derivatives) can dynamically adjust the parameter gradients according to the characteristics of the object. It will adjust the importance (gradient weight) of the angle parameter according to the aspect ratio. This mechanism can be vital for high-precision detection as a slight angle error would cause a serious accuracy drop for large aspect ratios objects. More importantly, we have proved that KLD is scale invariant. We further show that the KLD loss can be degenerated into the popular $l_{n}$-norm loss for horizontal detection. Experimental results on seven datasets using different detectors show its consistent superiority, and codes are available at https://github.com/yangxue0827/RotationDetection and https://github.com/open-mmlab/mmrotate.

1 Introduction

The paper argues that inherited regression losses limit high-precision rotated detection and proposes a deductive framework where horizontal detection is a special case of rotation detection. It uses KLD between Gaussian representations of bounding boxes to adapt parameter optimization and reports state-of-the-art results across seven datasets.

Motivation: Existing rotated detectors inherit horizontal-detection pipelines, but current regression losses struggle with large-aspect-ratio objects and other challenging scenes.These limitations produce disadvantages in high-precision detection.
Framework: The proposed framework develops rotation regression deductively, treating horizontal detection as a special case that recovers commonly used l_n-norm losses.This creates a unified regression framework for both detection settings.
Optimization mechanism: KLD dynamically adjusts parameter gradients according to object characteristics, including increasing the angle parameter’s importance for large-aspect-ratio objects.The motivation is that angle and center parameters have different importance for different object types.
KLD loss: KLD converts rotated bounding boxes into 2-D Gaussian distributions and measures prediction–ground-truth distance without independently comparing parameters with different scales and units.The formulation is presented as a clean and coherent regression loss.
Results: KLD is theoretically scale invariant, and experiments with two detectors on seven public datasets report new state-of-the-art rotation-detection performance.The paper identifies scale invariance as important for rotation detection and makes source code publicly available.

2 Background

Rotated detection methods largely adapt horizontal detection pipelines by adding angle prediction, but independently optimized parameters can limit high-precision performance. A deductive Gaussian-distance formulation instead seeks a unified rotation framework that includes horizontal detection as a special case.

Related works: Existing rotated detectors mainly adapt horizontal detectors with rotated boxes and extend l_n-norm or differentiable approximate IoU losses.This reflects an inductive design paradigm that starts from horizontal detection and generalizes it to rotation.
Inductive loss design: Horizontal regression predicts location and size parameters, while rotated regression adds an angle parameter handled for angular periodicity.The parameters represent box center, width, height, and rotation relative to anchor and target boxes.
Inductive loss design: Independent parameter optimization makes detection sensitive to under-fitting any parameter, with angle errors especially important for large-aspect-ratio objects.Different object types have different sensitivity to center, size, and angle parameters.
Deductive loss design: The deductive approach converts an arbitrary-oriented box into a 2-D Gaussian and computes a distance between Gaussian distributions as the loss.The representation uses a Gaussian mean and covariance derived from the rotated box.
Deductive loss design: GWD partially couples height, width, and angle but independently optimizes the center and is not scale invariant.Its center-distance and coupling terms improve high-precision detection, yet the remaining center behavior can shift results.
Deductive loss design: The proposed unified framework is designed so rotation detection can degenerate to commonly used horizontal regression loss in the horizontal case.This establishes horizontal detection as a special case rather than the starting point for constructing the rotation loss.

3 Proposed Approach

The proposed loss converts rotated boxes to Gaussian distributions and uses KLD to couple parameter optimization. Its gradients adapt to object geometry, while the formulation is scale invariant and retains a horizontal special case.

KLD regression loss: KLD computes the regression distance between Gaussian distributions representing the predicted and ground-truth rotated boxes.The detector decodes boxes, converts predictions and targets to Gaussians, and then calculates KLD without changing inference time.
Parameter coupling: Each KLD term creates chain coupling among center, size, and angle parameters, enabling joint rather than independent optimization.The parameter interactions make the optimization mechanism self-modulated.
Gradient analysis: KLD gradients adjust position optimization according to object scale and target orientation, increasing emphasis on offsets whose deviations sharply reduce IoU.Small objects or short edges receive greater attention along corresponding directions.
Gradient analysis: Larger aspect ratios cause KLD to assign greater optimization attention to angle, addressing the severe accuracy impact of slight angle errors.The analysis links angle-gradient weighting to the object's aspect ratio and describes the resulting parameter interaction as mutually reinforcing.
Scale invariance: KLD is affine invariant, including scale invariance, and is presented as suitable for replacing nondifferentiable rotated IoU loss.The invariance follows from transforming both Gaussian distributions by the same full-rank matrix.
Horizontal special case: For horizontal detection, the KLD formulation produces terms similar to the established horizontal regression loss, with differences in divisors for center terms.The paper treats horizontal detection as a special case of the general formulation.
Loss formulation: The final regression loss applies a nonlinear transformation and a hyperparameter to normalize and smooth the KLD distance within the detector's multitask objective.The paper mainly uses sqrt(D) or ln(D + 1), with τ modulating the overall regression loss.

4 Experiment

Experiments across multiple datasets and detectors evaluate KLD regression, showing gains in high-precision and broader rotated detection while retaining competitive horizontal detection performance.

Ablation Study: 9.17%, 1.58%, and 5.55% are KLD’s reported improvements on MLT, UCAS-AOD, and DOTA-v1.0, respectively.KLD also gains 3.63% and 3.53% on DOTA-v1.5 and DOTA-v2.0, which contain many tiny objects.
Peer Comparison: 6.07% is KLD’s AP75 improvement over R3Det, with a further 3.65% AP50:95 gain on DOTA-v1.0.RetinaNet-KLD also surpasses R3Det in AP50 and 7-AP50 under the stated comparison.
Horizontal Detection Verification: 0.6% is KLD’s improvement over other listed losses on RetinaNet for horizontal detection on MS COCO.The evaluation also reports that KLD is not worse on Faster R-CNN and FCOS.
State-of-the-Art Comparison: 80.63% is the reported DOTA accuracy achieved by the method with a large backbone and multi-scale testing.Single-scale RetinaNet-KLD-R50 and R3Det-KLD-R50 achieve 75.28% and 77.36%, respectively.

5 Discussions

The method has theoretical and experimental support but cannot be directly applied to quadrilateral detection. Its use may also extend to sensitive application fields.

The method cannot be directly applied to quadrilateral detection.
Potential applications include sensitive fields such as remote sensing, aviation, and unmanned aerial vehicles.
The approach achieves effectiveness across different detectors and datasets, according to the paper’s conclusion.

A.1 Proof of Scale Invariance of KLD

The proof shows that KLD between transformed Gaussian distributions is invariant under full-rank affine transformations, including scaling. The supplied figure reference concerns comparison against L2-norm and GWD under scaling.

Full-rank affine transformations preserve the KLD between the corresponding transformed Gaussian distributions.The transformation maps means and covariances as Mµ and MΣM⊤.
Figure 4 compares L2-norm, GWD, and KLD as the scaling factor varies.
Setting M = kI proves scale invariance of KLD.Here, I denotes the identity matrix.

A.2 Analysis of Dkl(Nt||Np)’s High-Precision Detection

The analysis explains how Dkl(Nt||Np) couples parameter gradients during optimization. Smaller target dimensions receive heavier matching penalties, while larger aspect ratios increase attention to angle optimization.

The parameter differences are defined as ∆x = xp − xt, ∆y = yp − yt, and ∆θ = θp − θt.
Figure 5 varies the targeted height while showing L2-norm, GWD, and KLD against parameters.
Under the stated assumption that other parameters are optimized, the position optimization is dynamically adjusted according to object scale.
Smaller targeted height or width produces a heavier penalty on the matching loss.The paper states that this favors higher matching precision for smaller dimensions.
Larger aspect ratios make the model pay more attention to angle optimization.The equality condition in the analyzed case is ht = wt.
Dkl(Nt||Np) has a similar gradient strategy to Dkl(Np||Nt), but couples parameters more tightly.

A.3 The Visualization of KLD’s Advantages

The visualization compares KLD with L2-norm and GWD to examine scale invariance and high-precision detection behavior. KLD remains invariant to scaling and emphasizes errors relevant to large-aspect-ratio objects.

The section compares KLD, L2-norm, and GWD to visualize scale invariance and high-precision detection advantages.
KLD remains unchanged as two boxes are enlarged by a scaling factor, unlike L2-norm and GWD.The comparison uses boxes whose dimensions vary with scaling factor s.
For large-aspect-ratio boxes, KLD more heavily penalizes shorter-edge length, center-position, and angle mismatches.These penalties correspond to quantities to which IoU is described as sensitive for such objects.