Source-linked AI summary
Hull First, Wake Second: Wake-Reliance Suppression for Robust Maritime Vessel Detection
Yefan Wang, Xingyu Wang, Ruibiao Zhu, Yusen Wu
TL;DR
Maritime vessel detectors can rely on elongated wakes instead of small or weak hulls, motivating a framework that separates hull evidence from wake context. HullWake uses hull-first detection, controlled directional wake extraction, and wake-reliance suppression objectives. On Curated-Wake, it improves detection performance and wake-oriented robustness relative to box-only and mask-supervised baselines.
Problem
Maritime detectors may rely on correlated wake context instead of verifying the vessel hull, creating a shortcut-learning problem under changing wake evidence and water clutter.
Method
HullWake separates proposal-centered hull evidence from directional wake context and regularizes wake influence with supervision, consistency, suppression, and decorrelation objectives.
Results
HullWake improves AP by 2.9 points and APNoWake by 6.0 points over Mask2Former, reduces FPWakeLike from 90 to 62, increases WG-AP by 5.7 points, and lowers ∆wake from 0.178 to 0.128.
Takeaways & Limitations
The results support evaluating maritime detectors for wake reliance using weak/no-wake robustness, wake-like false positives, worst-group AP, and confidence stability after wake attenuation.
Takeaways & Limitations
The oriented wake cue extractor avoids requiring ground-truth heading annotations.
Abstract
from arXiv · showhide
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.
1 Introduction
HullWake addresses wake reliance in maritime vessel detection by making hull evidence primary and wake context supportive, then evaluates robustness beyond overall AP.
- Maritime scenes contain unstable water context, while wakes can be easier to detect than small, low-contrast hulls.This mismatch can make wake context a shortcut rather than reliable vessel evidence.
- HullWake extracts proposal-centered hull features for the main detection path and directional wake context through bidirectional proposal-anchored corridors.Wake cues are explicitly controlled so they support rather than dominate detection.
- Wake-dominant predictions are suppressed with wake response supervision, wake-attenuated consistency, wake-only confidence suppression, and hull–wake decorrelation.
- The paper introduces a wake-oriented evaluation protocol that goes beyond overall AP.The protocol targets robustness to weak or absent wakes and wake-like clutter.
2 Related Work
Related work establishes maritime detection and wake modeling as useful but incomplete foundations for studying wake reliance as a shortcut-learning problem.
- Existing maritime datasets support ship, obstacle, and open-water object detection, but mainly report object-level detection or tracking performance.
- Ship wakes provide motion and directional cues and have been used for ship detection and motion analysis, typically as positive evidence.
- HullWake models wake explicitly while regularizing its influence so vessel detection remains focused on the hull.
- The study compares box-level detectors and mask-supervised segmentation models to test whether generic stronger supervision reduces wake reliance.
- Wake reliance is framed as maritime shortcut learning because wake cues can correlate with vessel labels and be weakened or intervened on at test time.
3 Method
HullWake decomposes proposal evidence into hull and wake components, using hull features for vessel detection and localization while controlling wake context as auxiliary evidence.
- Problem formulation: The framework decomposes each proposal into hull-oriented and wake-oriented representations under a hull-first, wake-second design.
- Problem formulation: The hull feature provides the main evidence for vessel existence and localization, while the wake feature supplies auxiliary context.
- Problem formulation: Wake reliance is defined by sharp prediction changes after weakening wake evidence or high confidence on wake-like water without a visible hull.
- Framework overview: The architecture combines a hull-dominant detector, oriented wake cue extractor, controlled fusion head, and wake-reliance losses.
- Hull-dominant base detector: Box regression remains on the hull feature so localization is tied to the vessel body rather than trailing water patterns.
3.3 Oriented Wake Cue Extractor
The oriented wake cue extractor predicts hull orientation and samples two opposite proposal-anchored corridors to capture directional wake context while suppressing off-axis texture. It also produces a soft wake response map supervised using wake and non-wake annotations.
- Orientation and corridors: Hull features predict a coarse orientation for each proposal without requiring ground-truth heading annotations.The orientation defines parallel and perpendicular directions for subsequent corridor construction.
- Orientation and corridors: Two candidate corridors extend in opposite directions along the predicted hull orientation to address bow–stern ambiguity.Each corridor is bounded by a learned length and width derived from the proposal scale.
- Directional aggregation: Directional attention samples and aggregates features along the corridors, favoring elongated trailing structures while suppressing off-axis texture.The two directions use opposite parallel offsets relative to the predicted orientation.
- Wake response supervision: The extractor predicts a soft wake response map supervised by annotated wake regions and non-wake masks for wake-like negatives and water clutter.These masks support hard-negative analysis in the curated dataset.
3.4 Controlled Hull–Wake Fusion
Controlled hull–wake fusion lets wake evidence adjust detector confidence while keeping vessel localization grounded in hull evidence. Wake therefore remains contextual rather than an independent decision source.
- Controlled fusion: Wake enters the detector through a bounded fusion head rather than serving as an independent decision source.The fusion uses a gated wake contribution to the detector representation.
- Controlled fusion: Wake can adjust confidence, but box regression remains predicted from the hull feature.This preserves hull-based localization even when wake context is incorporated.
3.5 Wake-Attenuated Consistency
Wake-attenuated consistency tests whether positive predictions remain stable when wake response is suppressed. The method constructs an attenuated proposal feature and penalizes prediction divergence before versus after attenuation.
- Attenuation construction: The method attenuates wake evidence in proposal features using a resized wake response and strength λatt ∈[0, 1].The attenuation strength controls how strongly wake response is removed at RoI resolution.
- Consistency objective: Positive proposals are trained to keep predictions close before and after wake attenuation.The consistency loss uses KL divergence between the original and attenuated prediction distributions.
- Consistency objective: The consistency objective compares original and attenuated predictions through an averaged KL-divergence loss over positive proposals.The lower path in Fig. 2 represents this consistency process.
3.6 Wake-Only Confidence Suppression
Wake-only confidence suppression uses separate hull-only and wake-only verifiers to enforce hull-dominant evidence. It also lowers wake-only confidence for negative proposals, especially wake-like clutter.
- Confidence verification: Two auxiliary verifiers measure hull-only and wake-only confidence separately.This separation enables direct supervision of the relative contribution of hull and wake evidence.
- Confidence verification: For positive proposals, hull-only confidence is required to exceed wake-only confidence by margin m.The margin establishes hull dominance during vessel verification.
- Negative suppression: For negative proposals, wake-only confidence should remain low, particularly around wake-like negatives and water-clutter regions.These cases target false vessel verification based solely on wake appearance.
- Negative suppression: The dominance loss prevents wake-only evidence from becoming sufficient for vessel verification.It combines the positive dominance requirement with negative-proposal suppression.
3.7 Hull–Wake Decorrelation
HullWake separates hull and wake descriptors and penalizes their linear dependence, while jointly supervising wake evidence and enforcing wake-reliance controls in the training objective.
- Hull–Wake Decorrelation: The normalized hull–wake decorrelation loss penalizes linear dependence between hull and wake descriptors without forcing wake cues to be ignored.The loss operates on mini-batch descriptor means and corresponds to the upper dashed path in Fig. 2.
- Wake Response Supervision: Wake response supervision uses annotated wake-region masks restricted to sampled corridors to localize actual wake evidence.Wake-like negatives and water clutter provide hard-negative regions rather than wake positives.
- Overall Objective: The overall objective combines base detection, wake-response, consistency, dominance, and decorrelation losses.The wake-response term is active in the main setting because curated-dataset wake-region masks are available.
4 Experiments
Experiments use Curated-Wake, added wake diagnostics, standard and wake-oriented metrics, and comparisons across box-level, mask-supervised, and HullWake models. HullWake improves overall and weak/no-wake performance while reducing wake-like false positives and wake dependence.
- Datasets and Diagnostic Protocol: Curated-Wake contains about 10,000 images from three public maritime sources, with reused or manually added vessel annotations and new wake diagnostics.Detection labels distinguish clear wake, weak/no wake, and ambiguous instances; segmentation labels cover hull, wake, wake-like negative, and water clutter regions.
- Datasets and Diagnostic Protocol: The evaluation reports AP, AP50, AP50:95, APNoWake, FPWakeLike, WG-AP, and ∆wake to measure ordinary detection and wake reliance.APNoWake targets weak/no-wake vessels, FPWakeLike counts false positives on wake-like water patterns, and ambiguous cases are excluded from group-wise AP.
- Implementation, Evaluation Metrics, and Results: Main comparisons include box-level detectors and mask-supervised segmentation baselines, with results measured on Curated-Wake using Faster R-CNN with ResNet-50-FPN as the default HullWake backbone.Main and ablation results are reported as mean±std over three random seeds.
- Implementation, Evaluation Metrics, and Results: Compared with Mask2Former, HullWake improves AP by 2.9 points and APNoWake by 6.0 points.These gains are reported under the same wake-oriented evaluation protocol.
- Implementation, Evaluation Metrics, and Results: HullWake reduces FPWakeLike from 90 to 62, increases WG-AP by 5.7 points, and lowers ∆wake from 0.178 to 0.128.The reported changes indicate improved wake-like false-positive control and confidence stability after wake attenuation.
- Implementation, Evaluation Metrics, and Results: Ablations show that naive wake fusion can harm weak/no-wake robustness, while wake response supervision, wake-only suppression, and the full model improve the robustness trade-off.The full model gives the best trade-off across APNoWake, FPWakeLike, WG-AP, and ∆wake.
- Implementation, Evaluation Metrics, and Results: Grad-CAM comparisons show baseline activations extending toward wake or wake-like water patterns, whereas HullWake produces more hull-centered activations.This qualitative diagnosis is consistent with lower FPWakeLike and smaller ∆wake.
5 Conclusion
The paper presents HullWake and Curated-Wake as a wake-oriented approach to maritime perception. Across comparisons with box-level and mask-supervised baselines, it reports stronger weak/no-wake robustness, fewer wake-like false positives, higher worst-group AP, and smaller confidence drops after wake attenuation.
- Conclusion: HullWake regularizes directional wake context so it supports rather than dominates hull-centered vessel verification.The method combines wake response supervision, wake-attenuated consistency, wake-only confidence suppression, and hull–wake decorrelation.
- Conclusion: Curated-Wake contains about 10,000 images from three public sources with added detection- and segmentation-level wake annotations.The dataset is introduced as a wake-oriented maritime resource.
- Conclusion: Experiments show improvements in overall AP, weak/no-wake robustness, wake-like false positives, worst-group AP, and confidence stability after wake attenuation.The conclusion contrasts these results with box-level detectors and mask-supervised segmentation baselines.
- Conclusion: The reported results indicate that generic detection or mask supervision alone is not sufficient to remove wake reliance.The conclusion states that wake evidence is useful when its shortcut effect is constrained by hull-centered verification.