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

Yefan Wang; Xingyu Wang; Ruibiao Zhu; Yusen Wu

cs.CVAug 27, 2026

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

Authors: Yefan Wang, Xingyu Wang, Ruibiao Zhu, Yusen Wu

Organizations: 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 · Fujian University of Technology, Fujian, China

Abstract

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.

Figures & tables

Figure 1: Hull-first and wake-second reasoning.

Figure 2: Overview of the proposed HullWake framework. The model follows a hull-first, wake-second design, where wake cues are explicitly extracted and regularized to support rather than dominate vessel detection.

Source dataset	Role	Selected images	Original annotation	Added labels
Ships/Vessels in Aerial Images [ 7 ]	Main source	$>$ 3k	Ship boxes	Wake attr., masks
SMD benchmark [ 13 , 16 , 15 ]	Maritime source	$>$ 3k	Object boxes	Wake attr., masks
SeaDronesSee [ 21 ]	Cross-view source	$>$ 3k	Boxes / tracks	Wake attr., masks
Curated dataset	Evaluation set	$\sim$ 10k	Reused or added hull boxes	Detection + segmentation

Table 1: Source datasets and added annotations.

Method	Backbone	AP	AP 50	AP 50:95	AP NoWake	FP ${}_{\mathrm{WakeLike}}\downarrow$	WG-AP	$\Delta_{\mathrm{wake}}\downarrow$
Faster R-CNN [ 17 ]	R50-FPN	$54.6{\pm}0.3$	$79.8{\pm}0.4$	$51.2{\pm}0.3$	$43.2{\pm}0.5$	$118{\pm}4$	$41.7{\pm}0.4$	$0.226{\pm}0.011$
Cascade R-CNN [ 2 ]	R50-FPN	$56.1{\pm}0.4$	$81.0{\pm}0.3$	$52.7{\pm}0.4$	$44.8{\pm}0.6$	$112{\pm}5$	$43.0{\pm}0.5$	$0.214{\pm}0.010$
RetinaNet [ 11 ]	R50-FPN	$51.9{\pm}0.5$	$77.2{\pm}0.5$	$48.5{\pm}0.4$	$40.6{\pm}0.7$	$131{\pm}6$	$39.4{\pm}0.6$	$0.239{\pm}0.013$
FCOS [ 20 ]	R50-FPN	$53.4{\pm}0.4$	$78.5{\pm}0.4$	$50.0{\pm}0.4$	$42.1{\pm}0.6$	$124{\pm}5$	$40.8{\pm}0.5$	$0.231{\pm}0.012$
YOLO11-m [ 8 ]	Default	$57.3{\pm}0.3$	$82.5{\pm}0.4$	$53.8{\pm}0.3$	$45.5{\pm}0.5$	$109{\pm}4$	$44.1{\pm}0.4$	$0.207{\pm}0.010$
LSKNet [ 9 ]	LSKNet-S	$58.1{\pm}0.4$	$83.3{\pm}0.5$	$54.6{\pm}0.4$	$46.7{\pm}0.6$	$102{\pm}5$	$45.4{\pm}0.5$	$0.198{\pm}0.010$

Table 2: Main results on Curated-Wake.

Figure 3: Examples of detection- and segmentation-level wake annotations.

Source	AP NoWake gain	FP red.	WG-AP gain
SMD	+10.2	46.8%	+9.3
SeaDronesSee	+8.7	43.5%	+7.8
Ships/Vessels	+10.5	48.1%	+9.6

Table 3: Source-wise robustness gains.

Figure 6: Grad-CAM diagnosis of wake reliance.

Variant	Wake Ext.	Wake Sup.	Cons.	Dom.	Dec.	AP NoWake	FP ${}_{\mathrm{WakeLike}}\downarrow$	WG-AP	$\Delta_{\mathrm{wake}}\downarrow$
Base detector	✗	✗	✗	✗	✗	$43.2{\pm}0.5$	$118{\pm}4$	$41.7{\pm}0.4$	$0.226{\pm}0.011$
Naive wake fusion	✓	✗	✗	✗	✗	$41.9{\pm}0.6$	$137{\pm}6$	$40.5{\pm}0.5$	$0.273{\pm}0.014$
+ wake response supervision	✓	✓	✗	✗	✗	$50.6{\pm}0.5$	$76{\pm}4$	$48.9{\pm}0.4$	$0.152{\pm}0.009$
+ wake-attenuated consistency	✓	✗	✓	✗	✗	$48.4{\pm}0.6$	$98{\pm}5$	$46.7{\pm}0.5$	$0.181{\pm}0.010$
+ wake-only suppression	✓	✗	✗	✓	✗	$49.7{\pm}0.5$	$82{\pm}4$	$47.9{\pm}0.4$	$0.166{\pm}0.009$
+ hull–wake decorrelation	✓	✗	✗	✗	✓	$49.1{\pm}0.5$	$91{\pm}5$	$47.2{\pm}0.5$	$0.158{\pm}0.008$

Table 4: Component ablation on Curated-Wake.

Ship Detection

See all 16 papers

HullProbe: Measuring Interventional Part Reliance in Maritime Vessel Classifiers
Oct 18, 2026Faisal Mohammad, Mohamed YounisImage Classification Feature Attribution
University of Maryland, Baltimore County · Johns Hopkins University Applied Physics Laboratory
WUTDet: A 100K-Scale Ship Detection Dataset and Benchmarks with Dense Small Objects
Apr 9, 2026Junxiong Liang, Mengwei Bao, Tianxiang Wang +3Ship Detection Maritime Transportation
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 +2
Strip Convolution and Direction-Aware Exclusion Loss for Oriented Ship Detection
Sep 23, 2026Bin Chen, Yuanyuan Liu, Peng Yang +1Remote Sensing Ship Detection
School of Information and Software Engineering, East China Jiaotong University, Nanchang 330013, China. · Jiangxi Vocational University of Foreign Studies, Nanchang 330099, China.

Oct 18, 2026J/K move · Enter open · S save

HullProbe: Measuring Interventional Part Reliance in Maritime Vessel Classifiers

Faisal Mohammad, Mohamed Younis

University of Maryland, Baltimore County · Johns Hopkins University Applied Physics Laboratory

Image Classification Feature Attribution Shortcut Learning Fine-Grained Image Classification Maritime Surveillance

Open paper AskPDF arXiv

Recognizing a vessel's type from an image supports maritime surveillance and sanctions monitoring, but a classifier can reach high accuracy by leaning on the generic hull instead of the more distinctive superstructure, and is then fooled when a vessel is re-hulled or occluded. De-biasing methods claim to remove such reliance, yet the field certifies them with worst-group accuracy, a proxy that needs group labels and only infers reliance from accuracy. We measure the reliance directly. HullProbe swaps a registered anatomical part of a vessel for another instance's and reads how much that part drives the decision, using a same-class swap to control for compositing artifacts. Because a surface vessel is a hull and a superstructure arranged along one axis, its parts form a 2x3 grid whose 2^6 = 64 swaps we evaluate exhaustively, and a single Walsh transform returns each part's interventional influence and its interaction spectrum. On planted shortcuts, where the true reliance is known, the measure tracks worst-group accuracy (r = -0.94 over 45 configurations), and replacing its learned segmenter with human annotations changes the hull-share by less than the audit's resolution. On the 26 vessel types of MARVEL, the classifier places about 46% of its corrected sensitivity on the generic hull yet follows the superstructure when the two conflict, and the group-label-free de-biasers available there do not detectably reduce the hull's share. The findings hold across three architectures and a second maritime dataset, and the method extends to fine-grained vehicle recognition on Stanford Cars.

Abstract

Ship Detection

Explore similar work

HullProbe: Measuring Interventional Part Reliance in Maritime Vessel Classifiers

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

Strip Convolution and Direction-Aware Exclusion Loss for Oriented Ship Detection