RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery

Huiran Sun

cs.CVMar 5, 2026

RMK RetinaNet: Rotated Multi-Kernel RetinaNet for Robust Oriented Object Detection in Remote Sensing Imagery

Authors: Huiran Sun

Organizations: School of Computer Science and Engineering, Changchun University of Technology, 2055 Yan’an Street, Chaoyang District, Changchun, 130012, Jilin Province, China

Abstract

Rotated object detection in remote sensing imagery is hindered by three major bottlenecks: non-adaptive receptive field utilization, inadequate long-range multi-scale feature fusion, and discontinuities in angle regression. To address these issues, we propose Rotated Multi-Kernel RetinaNet (RMK RetinaNet). First, we design a Multi-Scale Kernel (MSK) Block to strengthen adaptive multi-scale feature extraction. Second, we incorporate a Multi-Directional Contextual Anchor Attention (MDCAA) mechanism into the feature pyramid to enhance contextual modeling across scales and orientations. Third, we introduce a Bottom-up Path to preserve fine-grained spatial details that are often degraded during downsampling. Finally, we develop an Euler Angle Encoding Module (EAEM) to enable continuous and stable angle regression. Extensive experiments on DOTA-v1.0, HRSC2016, and UCAS-AOD show that RMK RetinaNet achieves performance comparable to state-of-the-art rotated object detectors while improving robustness in multi-scale and multi-orientation scenarios.

Figures & tables

Figure 1: Overview of RMK RetinaNet. The MSK Block consists of four layers, each constructed by the MSK Module (§ 3.2 ) . Only the bottommost MSK module does not perform downsampling; the remaining three modules all do.The outputs of the MSK Module are labeled as $M_{1}$ , $M_{2}$ , $M_{3}$ , and $M_{4}$ , which serve as inputs to both the MDCAA Module (§ 3.3 ) and the Bottom-up Path Module (§ 3.4 ) . $M_{1}$ , $M_{2}$ , $M_{3}$ , and $M_{4}$ are fed into the MDCAA Module to generate $CP_{2}$ , $CP_{3}$ , and $CP_{4}$ . Similarly, when fed into the Bottom-up Path, only the top-level output $N5$ is retained. Feature maps $CP_{4}$ , $N5$ , $CP_{3}$ , and $CP_{2}$ are then concatenated with the $C_{5}$ , $C_{4}$ , and $C_{3}$ feature maps from Rotation RetinaNet at corresponding scales. Finally, the combined features are processed by the Euler Angle Encoding Module (§ 3.5 ) to produce the final detection results.

Figure 2: (a) The MSK Block consists of four MSK Modules. Here, we use

l

to denote the

l

-th MSK Module (§ 3.2 ) , and the output of each MSK Module is denoted as

M_{l}

. (b) In the MDCAA Module (§ 3.3 ) , horizontal convolution is denoted as

H

Conv

, vertical convolution as

V

Conv

, convolution along the main diagonal direction as

Main

Diagonal

Conv

, and convolution along the anti-diagonal direction as

Anti

Diagonal

Conv

Model	Input Size	FLOPs	Params
Ours+PKI Module	1024 $\times$ 1024	65.693	33.612
Ours+MSK Module	1024 $\times$ 1024	65.693	33.558

Table 1: Computational cost analysis of our method with PKI and MSK modules on DOTA. FLOPs are measured with 1024

\times

1024 input resolution.

Figure 3: Visualization of detection results on the DOTA dataset, demonstrating the model’s performance on large-scale, obliquely oriented, and densely arranged objects. The examples include various object categories such as planes, large vehicles (LV), small vehicles (SV), tennis courts (TC), harbors, and swimming pools (SP).

Figure 4: Qualitative comparison of detection results on the DOTA dataset. Top row: Rotation RetinaNet exhibits wrong detections and missed detections; Bottom row: our RMK RetinaNet successfully detects all instances, including ships, large vehicles (LV), small vehicles (SV), and storage tanks (ST). See § 4.4 for details.

Different Settings of Rotated Multi-Kernel RetinaNet
MSK Module	$\checkmark$	$\times$	$\times$	$\checkmark$	$\checkmark$	$\checkmark$
MDCAA Module	$\times$	$\times$	$\times$	$\checkmark$	$\checkmark$	$\checkmark$
Bottom-up path Module	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$	$\checkmark$
Euler Angle Encoding Module	$\times$	$\times$	$\checkmark$	$\times$	$\times$	$\checkmark$
mAP	68.90	68.79	68.95	69.20	69.59	70.38

Table 2: Ablation study of key components in Rotated Multi-Kernel RetinaNet. We evaluate the contribution of each module by progressively adding MSK Module, MDCAA Module, Bottom-up Path Module, and Euler Angle Encoding Module. The results show that the full combination achieves the best mAP of 70.38%.

Figure 5: Performance comparison of object detection on the HRSC2016 dataset. We divide the HRSC2016 dataset into 4 categories, and our method achieves 68.77% mAP, significantly outperforming the baseline model.

Methods	Backbone	Plane	BD	Bridge	GTF	SV	LV	Ship	TC	BC	ST	SBF	RA	Harbor	SP	HC	mAP
R-FCN [ 6 ]	ResNet-101	81.01	58.96	31.64	58.97	49.77	45.04	49.29	68.99	52.07	67.42	41.83	51.44	45.15	53.30	33.89	52.58
IENet [ 19 ]	ResNet-101	80.20	62.54	39.82	32.07	49.71	65.01	52.58	81.45	44.66	78.51	46.54	56.73	64.40	64.24	36.75	57.14
R-DFPN [ 38 ]	ResNet-101	80.92	65.82	33.77	58.94	55.77	50.94	54.78	90.33	66.34	68.66	48.73	51.76	55.10	51.32	35.88	57.94
PIoU [ 4 ]	DLA-34	80.9	69.70	24.10	60.20	38.30	64.40	64.80	90.90	77.20	70.40	46.50	37.10	57.10	61.90	64.00	60.50
R 2 CNN [ 15 ]	ResNet-101	80.94	65.67	35.34	67.44	59.92	50.91	55.81	90.67	66.92	72.39	55.06	52.23	55.14	53.35	48.22	60.67
RRPN [ 23 ]	ResNet-101	88.52	71.20	31.66	59.30	51.85	56.19	57.25	90.81	72.84	67.38	56.69	52.84	53.08	51.94	53.58	61.01

Table 3: Detection accuracy comparison on DOTA dataset. For each method, we report Average Precision (AP) for 15 object categories: Plane, Baseball Diamond (BD), Bridge, Ground Track Field (GTF), Small Vehicle (SV), Large Vehicle (LV), Ship, Tennis Court (TC), Basketball Court (BC), Storage Tank (ST), Soccer Ball Field (SBF), Roundabout (RA), Harbor, Swimming Pool (SP), and Helicopter (HC). The overall mean AP (mAP) is also provided. Our method achieves the best performance with 70.38% mAP, demonstrating significant improvements over previous approaches.

Method	mAP	Plane	Car
YOLOv2 [ 26 ]	87.90	96.60	79.20
R-DFPN [ 38 ]	89.20	95.90	82.50
DRBox [ 20 ]	89.95	94.90	85.00
Rotation RetinaNet [ 43 ]	91.495	96.12	86.87
Ours	91.735	97.01	86.46

Table 4: Performance evaluation on the UCAS-AOD dataset. We compare our method against YOLOv2, R-DFPN, DRBox, and Rotation RetinaNet. Results are reported in terms of mAP and per-class AP for Plane and Car. Our approach achieves the best overall performance with 91.735% mAP.

Explore similar work

Fourier Angle Alignment for Oriented Object Detection in Remote Sensing
Feb 27, 2026Changyu Gu, Linwei Chen, Lin Gu +1Object Detection Aerial Object Detection
Beijing Institute of Technology · University of Hong Kong · The Hong Kong University of Science and Technology +2
Real-Time Oriented Object Detection Transformer in Remote Sensing Images
Mar 16, 2026Zeyu Ding, Yong Zhou, Jiaqi Zhao +4Object Detection Detection Transformer
School of Computer Science and Technology / School of Artificial Intelligence, China University of Mining and Technology, Xuzhou 221116, China · Mine Digitization Engineering Research Center of the Ministry of Education, Xuzhou 221116, China · Jiangsu Provincial Industrial Technology Engineering Center for Intelligent Sensing and Emergency IoT in Underground Space, Xuzhou 221116, China +1
FGAA-FPN: Foreground-Guided Angle-Aware Feature Pyramid Network for Oriented Object Detection
Feb 11, 2026Jialin MaMulti-Scale Feature Fusion Object Detection
International school of BUPT, Beijing University Of Posts And Telecommunications, Beijing,China.

Feb 27, 2026J/K move · Enter open · S save

Fourier Angle Alignment for Oriented Object Detection in Remote Sensing

Changyu Gu, Linwei Chen, Lin Gu, Ying Fu

Beijing Institute of Technology · University of Hong Kong · The Hong Kong University of Science and Technology+2

Object Detection Aerial Object Detection

Open paper AskPDF arXiv

In remote sensing rotated object detection, mainstream methods suffer from two bottlenecks, directional incoherence at detector neck and task conflict at detecting head. Ulitising fourier rotation equivariance, we introduce Fourier Angle Alignment, which analyses angle information through frequency spectrum and aligns the main direction to a certain orientation. Then we propose two plug and play modules : FAAFusion and FAA Head. FAAFusion works at the detector neck, aligning the main direction of higher-level features to the lower-level features and then fusing them. FAA Head serves as a new detection head, which pre-aligns RoI features to a canonical angle and adds them to the original features before classification and regression. Experiments on DOTA-v1.0, DOTA-v1.5 and HRSC2016 show that our method can greatly improve previous work. Particularly, our method achieves new state-of-the-art results of 78.72% mAP on DOTA-v1.0 and 72.28% mAP on DOTA-v1.5 datasets with single scale training and testing, validating the efficacy of our approach in remote sensing object detection. The code is made publicly available at https://github.com/gcy0423/Fourier-Angle-Alignment .

Abstract

Explore similar work

Fourier Angle Alignment for Oriented Object Detection in Remote Sensing

Real-Time Oriented Object Detection Transformer in Remote Sensing Images

FGAA-FPN: Foreground-Guided Angle-Aware Feature Pyramid Network for Oriented Object Detection