WUTDet: A 100K-Scale Ship Detection Dataset and Benchmarks with Dense Small Objects

Junxiong Liang; Mengwei Bao; Tianxiang Wang; Xinggang Wang; An-An Liu; Ryan Wen Liu

cs.CVApr 9, 2026

WUTDet: A 100K-Scale Ship Detection Dataset and Benchmarks with Dense Small Objects

Authors: Junxiong Liang, Mengwei Bao, Tianxiang Wang, Xinggang Wang, An-An Liu, Ryan Wen Liu

Organizations: School of Navigation, Wuhan University of Technology, Wuhan 430063, China · Sanya Science and Education Innovation Park, Wuhan University of Technology, Sanya 572000, China · State Key Laboratory of Maritime Technology and Safety, Wuhan 430063, China · School of Electronic Information and Communications, Huazhong University of Science and Technology, Wuhan 430074, China ·

Abstract

Ship detection for navigation is a fundamental perception task in intelligent waterway transportation systems. However, existing public ship detection datasets remain limited in terms of scale, the proportion of small-object instances, and scene diversity, which hinders the systematic evaluation and generalization study of detection algorithms in complex maritime environments. To this end, we construct WUTDet, a large-scale ship detection dataset. WUTDet contains 100,576 images and 381,378 annotated ship instances, covering diverse operational scenarios such as ports, anchorages, navigation, and berthing, as well as various imaging conditions including fog, glare, low-lightness, and rain, thereby exhibiting substantial diversity and challenge. Based on WUTDet, we systematically evaluate 20 baseline models from three mainstream detection architectures, namely CNN, Transformer, and Mamba. Experimental results show that the Transformer architecture achieves superior overall detection accuracy (AP) and small-object detection performance (APs), demonstrating stronger adaptability to complex maritime scenes; the CNN architecture maintains an advantage in inference efficiency, making it more suitable for real-time applications; and the Mamba architecture achieves a favorable balance between detection accuracy and computational efficiency. Furthermore, we construct a unified cross-dataset test set, Ship-GEN, to evaluate model generalization. Results on Ship-GEN show that models trained on WUTDet exhibit stronger generalization under different data distributions. These findings demonstrate that WUTDet provides effective data support for the research, evaluation, and generalization analysis of ship detection algorithms in complex maritime scenarios. The dataset is publicly available at: https://github.com/MAPGroup/WUTDet.

Figures & tables

Fig. 1: Overview of the proposed framework, including the construction of the large-scale ship detection dataset (WUTDet), the systematic evaluation of detection models with different architectures, and the construction and validation of the cross-dataset test set (Ship-GEN). WUTDet is characterized by dense small objects and diverse scenarios, while Ship-GEN is used to validate the effectiveness of WUTDet and the generalization capability of the models.

Dataset	Year	Image/Frame	Object	Size	Category	Object Size
Dataset	Year	Image/Frame	Object	Size	Category	$0\!\sim\!32^{2}$ (Small)		$32^{2}\!\sim\!96^{2}$ (Medium)		$>96^{2}$ (Large)
SMD [ 20 ]	2017	20,367	157,998	1920 $\times$ 1080	10	16,017	10.14%	77,263	48.90%	6,4718	40.96%
Seaships † [ 17 ]	2018	31,455	40,077	1920 $\times$ 1080	6	-	-	-	-	-	-
SeaShips(7000) [ 17 ]	2018	7,000	9,221	1920 $\times$ 1080	6	15	0.16%	784	8.50%	8,422	91.33%
Mcships † [ 21 ]	2020	14,709	26,529	random	13	-	-	-	-	-	-
Mcships-lite [ 21 ]	2020	7,996	11,331	random	2	949	8.38%	2,038	17.99%	8,344	73.64%

TABLE I: Overview of Ship Detection–Related Datasets. This table provides a statistical comparison of different datasets in terms of release year, number of images, number of annotated objects, image resolution, annotation categories, and the quantity and proportion of targets at different scales. Here, ”

\dagger

” indicates that the corresponding dataset is unavailable, and “random” denotes that the dataset contains images with more than two different resolutions, with most samples having a resolution lower than 1920

\times

1080. Bold indicate the highest values in the corresponding columns.

Fig. 2: WUTDet Data Collection Setup. (a) Onboard data acquisition device used for image collection; (b) illustration of the ship navigation route over three months during data collection process.

Fig. 3: Image examples in WUTDet from four typical operational scenarios, namely port, anchorage, navigation, and berthing. These scenarios differ significantly in ship distribution, target scale, and occlusion level, posing diverse challenges for ship detection.

Split	1920 $\times$ 1080	2560 $\times$ 1440	Images	Percentage	Objects	Category
Train	65,070	15,391	80,461	80.0%	305,264	Ship
Validation	8,134	1,924	10,058	10.0%	38,118	Ship
Test	8,133	1,924	10,057	10.0%	37,996	Ship
Total	81,337	19,239	100,576	100%	381,378	Ship

TABLE II: Train/Validation/Test Split Statistics of WUTDet. Considering that ship type is of limited importance for collision avoidance in autonomous navigation and that small-scale objects do not provide sufficient visual semantic information for reliable fine-grained annotation, all ship instances in WUTDet are uniformly annotated as a single category.

Fig. 4: Image examples in WUTDet under fog, glare, low-lightness, rain, and normal conditions. These diverse imaging conditions provide data support for evaluating the performance of different models under heterogeneous visual conditions.

Scene	Fog	Glare	Low-lightness	Rain	Normal
Images	3,380	900	578	694	14,688
Objects	9,688	2,995	2,294	2,125	59,183

TABLE III: Image and object statistics under different scene conditions in WUTDet.

Fig. 5: Presents the statistics of object scales in WUTDet. (a) shows the number and proportion of objects in different area intervals, where Extremely Small, Relatively Small, Generally Small, Medium-Small, Medium-Large, and Large correspond to area ranges of

[0,12^{2}]

(12^{2},20^{2}]

(20^{2},32^{2}]

(32^{2},60^{2}]

(60^{2},96^{2}]

, and

(96^{2},+\infty)

pixels, respectively. (b) illustrates the frequency distribution of object areas.

Fig. 6: Distribution of the number of objects per image in WUTDet. The bars show the statistics of the training, validation, and test sets, while the red curve denotes a smoothed approximation of the object-count distribution in the training set.

Fig. 7: Aspect ratio distribution of objects in WUTDet. The aspect ratio is defined as the height-to-width ratio of the bounding box. The histogram shows the frequency distribution of aspect ratios, and the orange curve indicates the corresponding smoothed trend.

Fig. 8: Pipeline of the YOLOv8 architecture [ 41 ] . As a representative CNN-based object detection model, YOLOv8 improves detection efficiency and representation capability through designs such as an anchor-free decoupled head and enhanced gradient flow, and has become a widely adopted baseline for real-time object detection tasks.

Fig. 9: Pipeline of the RT-DETR architecture [ 14 ] . As a representative Transformer-based object detection model, RT-DETR [ 14 ] improves computational efficiency and spatial alignment through designs such as the efficient hybrid encoder and IoU-aware query selection, and achieves end-to-end real-time detection without post-processing by eliminating NMS.

Fig. 10: Pipeline of the Mamba-YOLO architecture [ 40 ] . As a representative Mamba-based object detection model, Mamba-YOLO [ 40 ] introduces SSM to efficiently capture long-range dependencies with linear computational complexity. By avoiding the quadratic overhead of traditional attention mechanisms, it achieves a favorable balance between global contextual modeling and real-time inference efficiency.

Category	Method	Backbone	Year	Params(M)	Epoch	Input size	FPS $\uparrow$	AR 50:95 (%) $\uparrow$	AP 50:95 (%) $\uparrow$	AP s (%) $\uparrow$	AP m (%) $\uparrow$	AP l (%) $\uparrow$
CNN	CenterNet [ 51 ]	Resdcn-18	2019	14.4	140	$640\times 640$	38.0	61.3	54.2	31.1	58.1	80.9
	YOLOX-S [ 33 ]	Modified CSP	2021	8.9	300	$640\times 640$	60.8	70.9	67.6	50.0	69.5	86.0
	YOLOv6-S [ 42 ]	EfficientRep	2022	18.5	400	$640\times 640$	33.08	67.6	64.0	36.8	69.8	87.2
	YOLOv8-S [ 41 ]	C2f	2023	11.1	300	$640\times 640$	60.1	66.2	62.5	33.5	69.1	87.2
	YOLOv11-S [ 41 ]	C3K2	2024	9.4	300	$640\times 640$	57.0	66.4	62.6	33.9	69.2	87.1
	Hyper-YOLO-S [ 43 ]	MANet	2024	14.8	300	$640\times 640$	66.7	29.4	63.1	34.5	69.6	87.3

TABLE IV: Performance of different object detection methods on our dataset. Since the authors of YOLOv8-S, YOLOv11-S, Hyper-YOLO-S, YOLOv12-S, YOLOv13-S, and FBRT-YOLO-S do not explicitly specify the module definitions of their backbones, the core structures proposed at the backbone stage in each method are uniformly regarded and labeled as the backbone network in this work. FPS is measured with a batch size of 16.

\uparrow

indicates that higher values correspond to better performance. Bold indicates the best result, and underlining indicates the second-best result.

Fig. 11: Visualization of detection results of different CNN-based object detection methods on WUTDet. CT denotes the prediction confidence score, whose value is set to the default threshold used for model visualization. The time displayed in the figure does not represent the actual image acquisition time. (a1-a4) represent CenterNet CT = 0.30; (b1-b4) represent YOLOX-S CT = 0.25; (c1-c4) represent YOLOv6-S CT = 0.40; (d1-d4) represent YOLOv8-S CT = 0.25; (e1-e4) represent YOLOv11-S CT = 0.25; (f1-f4) represent Hyper-YOLO-S CT = 0.25; (g1-g4) represent YOLOv12-S CT = 0.25; (h1-h4) represent YOLOv13-S CT = 0.25; (i1-i4) represent FBRT-YOLO-S CT = 0.25; (j1-j4) represent YOLO-MS-S CT = 0.30.

Fig. 12: Continuation of Fig. 11 . Visualization of detection results of different Transformer- and Mamba-based object detection methods on WUTDet. CT denotes the prediction confidence score, whose value is set to the default threshold used for model visualization. (k1-k4) represent Deformable DETR CT = 0.30; (l1-l4) represent DAB-DETR-R50 CT = 0.30; (m1-m4) represent DINO-4scale CT = 0.30; (n1-n4) represent RT-DETR-R50VD-M CT = 0.60; (o1-o4) represent D-FINE-S CT = 0.40; (p1-p4) represent LW-DETR-S CT = 0.50; (q1-q4) represent DEIM-D-FINE-S CT = 0.50; (r1-r4) represent VMamba-T-FasterRcnn@1x CT = 0.30; (s1-s4) represent VMamba-T-FasterRcnn@1x CT = 0.30; (t1-t4) represent MobileMamba-B1-RetinaNet@1x CT = 0.30; (u1-u4) represent MobileMamba-B1-RetinaNet@1x CT = 0.30; (v1-v4) represent Mamba YOLO-M CT = 0.25.

Category	Method	Input size	AP 50:95 (%) $\uparrow$
Category	Method	Input size	Fog	Glare	Low-lightness	Rain	Normal
CNN	CenterNet [ 51 ]	$640\times 640$	58.2	53.7	61.7	45.4	53.6
	YOLOX-S [ 33 ]	$640\times 640$	69.9	66.6	69.8	60.5	67.3
	YOLOv6-S [ 42 ]	$640\times 640$	68.3	65.3	70.4	58.0	63.1
	YOLOv8-S [ 41 ]	$640\times 640$	67.1	63.9	69.9	56.7	61.7
	YOLOv11-S [ 41 ]	$640\times 640$	67.3	63.9	69.9	56.7	61.8

TABLE V: Performance of different object detection methods on our dataset under different scenarios.

\uparrow

indicates that higher values correspond to better performance. Bold indicates the best result, and underlining indicates the second-best result.

Fig. 13: Visualization of detection results of different CNN-based object detection methods on WUTDet under various weather and illumination conditions. CT denotes the prediction confidence score, whose value is set to the default threshold used for model visualization. (a1-a5) represent CenterNet CT = 0.30; (b1-b5) represent YOLOX-S CT = 0.25; (c1-c5) represent YOLOv6-S CT = 0.40; (d1-d5) represent YOLOv8-S CT = 0.25; (e1-e5) represent YOLOv11-S CT = 0.25; (f1-f5) represent Hyper-YOLO-S CT = 0.25; (g1-g5) represent YOLOv12-S CT = 0.25; (h1-h5) represent YOLOv13-S CT = 0.25; (i1-i5) represent FBRT-YOLO-S CT = 0.25; (j1-j5) represent YOLO-MS-S CT = 0.30.

Fig. 14: Continuation of Fig. 13 . Visualization of detection results of different Transformer- and Mamba-based object detection methods on WUTDet under various weather and illumination conditions. CT denotes the prediction confidence score, whose value is set to the default threshold used for model visualization. (k1-k5) represent Deformable DETR CT = 0.30; (l1-l5) represent DAB-DETR-R50 CT = 0.30; (m1-m5) represent DINO-4scale CT = 0.30; (n1-n5) represent RT-DETR-R50VD-M CT = 0.60; (o1-o5) represent D-FINE-S CT = 0.40; (p1-p5) represent LW-DETR-S CT = 0.50; (q1-q5) represent DEIM-D-FINE-S CT = 0.50; (r1-r5) represent VMamba-T-FasterRcnn@1x CT = 0.30; (s1-s5) represent VMamba-T-FasterRcnn@1x CT = 0.30; (t1-t5) represent MobileMamba-B1-RetinaNet@1x CT = 0.30; (u1-u5) represent MobileMamba-B1-RetinaNet@1x CT = 0.30; (v1-v5) represent Mamba YOLO-M CT = 0.25.

Method	Training Set	Testing Set	AR 50:95 (%) $\uparrow$	AP 50:95 (%) $\uparrow$	AP S (%) $\uparrow$	AP M (%) $\uparrow$	AP L (%) $\uparrow$
Category: CNN
Hyper-YOLO-S [ 43 ]	WSODD	Ship-GEN	34.7	26.9	3.5	23.5	45.1
	WaterScence	Ship-GEN	29.7	23.6	6.8	17.0	41.0
	WUTDet	Ship-GEN	46.2	38.8	25.9	43.0	41.3
YOLOv6-S [ 42 ]	WSODD	Ship-GEN	34.9	25.8	3.4	23.4	42.3
	WaterScence	Ship-GEN	33.6	25.3	7.3	18.2	44.1

TABLE VI: Performance of different object detection methods trained on different datasets and evaluated on the Ship-GEN dataset.Note that the input size of MobileMamba-B1-RetinaNet@1x is set to 640

\times

640.

\uparrow

indicates that higher values correspond to better performance. Purple highlights results obtained on WUTDet, and bold indicates the best result.

Ship Detection

See all 16 papers

Dec 10, 2025

LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing

General-purpose object detectors face fundamental structural limitations when applied to ship detection in satellite imagery, where the ship scale distribution is concentrated at small sizes and high aspect ratios. In conventional You Only Look Once architectures, the deepest feature pyramid level (stride 32) compresses narrow vessels into sub-pixel representations, causing severe spatial feature dilution and compromising accurate ship boundary regression. We propose Less is More YOLO, a streamlined detector built upon the extra-large variant of YOLOv9, to address these domain-specific structural conflicts. From a statistical analysis of ship scale distributions across four major benchmarks (SODA-A, DOTA-v1.5, FAIR1M-v2.0, and ShipRSImageNet), we introduce a Pyramid Level Shift Strategy that shifts the detection head from strides 8, 16, and 32 to strides 4, 8, and 16. This shift satisfies a spatial representability condition derived from the Nyquist-Shannon principle for the narrowest targets, while eliminating the computational redundancy of the deepest pyramid level. To further stabilize training on high-resolution satellite inputs, we incorporate a group-normalized composite-backbone projection module, mitigating gradient instability in memory-constrained micro-batch regimes. Validated on these four datasets, our detector attains an mAP50:95 of 0.600 with only 21.16 million parameters, a 64.1% reduction from the extra-large YOLOv9 baseline (58.99 million). Despite this compact size, our model surpasses state-of-the-art detectors up to three times larger, validating that a well-targeted pyramid level shift achieves a "Less is More" balance between accuracy and efficiency. The code is available at https://github.com/egshkim/LiM-YOLO.

Seon-Hoon Kim, Yerin Kim, Hyeji Sim +3

University of Science and Technology (UST), Daejeon 34113, Republic of Korea · Korea Aerospace Research Institute (KARI), Daejeon 34133, Republic of Korea

Remote Sensing Small Object Detection Small Object Detection Efficient Object Detection Ship Detection You Only Look Once Object Detection+1

Ask PDF

Aug 3, 2026

OSSDD - a New Open Dataset for Sentinel-1 Ship Detection

Ship detection in Synthetic Aperture Radar (SAR) images plays an important role for maritime situational awareness, especially with respect to different illegal activities at sea such as illegal fishing, smuggling or border violations. Modern ship detection methods using neural networks usually require large training datasets, which are considerably scarcer in the SAR domain than in the electro-optical domain. While several free datasets exist for this task, their availability and usability vary. In this paper, OpenSARShip-Ship Detection Dataset (OSSDD), a new dataset based on the well-known OpenSARShip 1.0 dataset is proposed for training neural networks for SAR ship detection. OSSDD is freely available and contains 15,197 Sentinel-1 amplitude patches in VV and VH polarization, binary ship masks, axis-aligned bounding box and rotated bounding box annotations for a total of 55,759 ships. The construction of the dataset, the contents and structure of the downloadable data and experiments with three common detector models (Faster R-CNN, FCOS, DETR) are shown and discussed. The results serve as benchmarks for future experiments. The dataset is available on Hugging Face at https://huggingface.co/datasets/sylviaHoch/OpenSARShip-Ship-Detection-Dataset.

Horst Hammer, Sylvia Hochstuhl, Antje Thiele +4

Fraunhofer Institute of Optronics, System Technologies and Image Exploitation - IOSB, Gutleuthausstraße 1, Ettlingen, 76275, Germany · Karlsruhe Institute of Technology - KIT, Kaiserstraße 12, Karlsruhe, 76131, Germany · HENSOLDT Sensors GmbH, Wörthstraße 85, Ulm, 89077, Germany

Object Detection Synthetic Aperture Radar Remote Sensing Benchmarks Ship Detection

Ask PDF

Aug 27, 2026

Hull First, Wake Second: Wake-Reliance Suppression for Robust Maritime Vessel Detection

Maritime vessel detectors often face scenes where hulls are small, low-contrast, or blurred, while wakes are longer and easier to detect. This creates a wake-reliance problem: detectors may miss slow or stationary vessels with weak wakes, or produce false positives on wake-like water clutter. We propose HullWake, a hull-first wake-second framework for robust maritime vessel detection. HullWake separates proposal-centered hull evidence from directional wake context, extracts wake cues with bidirectional proposal-anchored corridors, and suppresses wake-dominant predictions through wake response supervision, wake-attenuated consistency, wake-only confidence suppression, and hull--wake decorrelation. We also introduce a wake-oriented evaluation protocol covering weak/no-wake vessels, wake-like hard negatives, worst-group AP, and confidence drop after wake attenuation. Experiments are conducted on Curated-Wake, a wake-oriented maritime dataset of about 10,000 images curated from Ships/Vessels in Aerial Images, the SMD benchmark, and SeaDronesSee, with newly added detection- and segmentation-level wake annotations. Compared with box-only detectors and mask-supervised segmentation baselines, HullWake improves overall AP, weak/no-wake robustness, wake-like false positives, worst-group AP, and confidence stability after wake attenuation.

Yefan Wang, Xingyu Wang, Ruibiao Zhu +1

University of Shanghai for Science and Technology, Shanghai, China · University of Science and Technology Liaoning, Liaoning, China · School of Computing, College of Systems and Society, The Australian National University, ACT, Australia +1

Ship Detection Object Detection Spurious Correlation Robustness Aerial Object Detection

Ask PDF

Abstract

Ship Detection

Explore similar work

LiM-YOLO: Less is More with Pyramid Level Shift for Ship Detection in Optical Remote Sensing

OSSDD - a New Open Dataset for Sentinel-1 Ship Detection

Hull First, Wake Second: Wake-Reliance Suppression for Robust Maritime Vessel Detection