{"id":627,"date":"2026-08-30T20:52:09","date_gmt":"2026-08-30T18:52:09","guid":{"rendered":"https:\/\/vap.aau.dk\/marinevision\/?page_id=627"},"modified":"2026-09-04T20:06:33","modified_gmt":"2026-09-04T18:06:33","slug":"short-papers","status":"publish","type":"page","link":"https:\/\/vap.aau.dk\/marinevision\/short-papers\/","title":{"rendered":"Short papers"},"content":{"rendered":"\n
\n\n\n\n

This is the list of short papers accepted for presentation at the 2nd Workshop on Marine Vision 2026<\/strong>. The papers are publicly available through OpenReview. The papers appear in random order.<\/p>\n\n\n\n

\n\n\n\n
\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Semantic Late Interaction for Cross-Temporal Reef Relocalization<\/a><\/h4>\n\n\n\n

Hugues Sibille, Jonathan Sauder, Guilhem Banc-Prandi, Devis Tuia<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Visual relocalization in coral reefs is hard due to changes in water conditions, lighting and the scene itself between revisits. For practical reef monitoring, relocalizing a place along a transect across passes remains an open question. We introduce a cross-temporal reef visual relocalization benchmark, with clear difficulty tiers that isolate reverse viewpoints and long-term scene change. On this benchmark, we evaluate a range of relocalization methods across local and global descriptors and assess the impact of in-domain learned semantic information. We find that matching ALIKED keypoints described by a Coralscapes-fine-tuned DINOv3 encoder, and scoring frame pairs by a ColBERT-style late interaction mechanism, outperforms off-the-shelf feature descriptors and a strong general-purpose global descriptor (AnyLoc). The improvement is carried by domain-tuned semantics: reef-tuned per-keypoint descriptors survive the appearance change between passes where geometric-keypoint descriptors do not.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Marine Wildlife Individual Re-identification Using Deep Learning: A Component Variation Study with Picasso Triggerfish<\/a><\/h4>\n\n\n\n

Tedi Yankov, Niki Amini-Naieni, Oliver N. F. King, Andrew Zisserman, Cait Newport<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Reliable individual wildlife re-identification (re-ID) is essential for ecological monitoring but remains particularly challenging in underwater environments. Although modern re-ID pipelines comprise multiple stages – detection\/segmentation, feature extraction, and representation learning – the impact of these design choices has not been systematically evaluated end-to-end. We present a component variation study of an underwater wildlife re-ID pipeline, comparing 132 configurations across eleven detection and segmentation methods, four feature representation backbones, and three learning objectives, evaluated on a unique dataset of 37 individual Picasso triggerfish (Rhinecanthus aculeatus) in natural reef and controlled laboratory settings. Our results demonstrate that re-ID is now tractable in controlled and simple in-situ conditions: the best latency-constrained pipeline achieves a composite open-set score of 0.962 at
\u22481 ms per image. We identify backbone choice as the dominant performance driver, analyse accuracy-latency trade-offs, and provide practical guidance for pipeline design.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Classifying Ships via Underwater Acoustic Spectrograms in the UniqueShip Dataset<\/a><\/h4>\n\n\n\n

Connor Hashemi, Trevor Stout, Jason Parham, Anthony Hoogs<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Underwater Acoustic Target Recognition (UATR) enables the passive identification and classification of distant marine vessels by their unique acoustic signatures. However, the development of machine learning models for UATR is limited by the small number of large, diverse, and publicly available labeled datasets. In this paper, we introduce UniqueShip, a machine learning-ready benchmark dataset for UATR applications sourced from the open Ocean Networks Canada (ONC) repository. By converting 5-second audio recordings into spectrograms that capture their time-frequency content, we train image-based models to identify ship types. Initial benchmark evaluations indicate 66.5% accuracy using transformer-based models, with Mel spectrograms outperforming standard Short-Time Fourier Transform (STFT). The dataset is provided at uniqueshipdata.org.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Bootstrapped Watershed: Towards Few-Shot Instance ZooScan Segmentation<\/a><\/h4>\n\n\n\n

Pratham Tatraiya, Torben Globisch, Vivian Fischbach, Stefan Oehmcke<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

ZooScan can digitize plankton samples, but extracting individual organisms depends on manually annotated training data. This becomes limiting when a single scan contains hundreds or thousands of touching organisms. We address this problem with Bootstrapped Watershed, a few-shot instance-segmentation pipeline. Using frozen DINOv3 features, we train a shallow MLP on three annotated crops and use it to produce pseudo-labels for additional crops. These pseudo-labels are used to train a U-Net decoder attached to a frozen ConvNeXt-Small encoder. For full-scan inference, Gaussian blending combines overlapping tile predictions, after which boundary removal, pinch-point severing, and distance-transform watershed recover individual instances. The proposed pipeline reaches a panoptic quality of 0.310 and instance recall of 0.573. Direct U-Net training reaches 0.273 and 0.538, respectively, while the best zero-training baseline reaches 0.171 and 0.202.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Gaussian-Splat Synthetic Augmentation for Underwater Anomaly Detection: A Substitution Feasibility Study<\/a><\/h4>\n\n\n\n

Jan Mikolajczyk, Maciej Malewicz, Aleksandra Lipi\u0144ska, Jan Matusiak<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Underwater anomaly detection is constrained by scarce site-specific training data and optical domain shifts between survey sessions. Physics-aware Gaussian Splatting decouples scene geometry from the scattering medium, enabling site-specific synthetic data. We ask whether Gaussian-Splat renders can replace real normal training images and how much substitution is tolerable before detection degrades. Reconstructing a turbid freshwater lake with WaterSplatting, we sweep from all-real to all-synthetic training at a fixed 50-image budget on two detectors: PatchCore (memory-based, training-free) and Reverse Distillation (feature-learning). Both hold performance to a
\u223c60% synthetic plateau; Reverse Distillation reaches AUPR 0.87 \/ AUROC 0.94 with 30 of 50 training images synthetic. Full substitution remains usable but costs AUPR
\u221217.6 % (AUROC \u22128.4 %), attributable to a photometric appearance gap. Gaussian-Splat renders are thus a viable partial substitute for real training data in data-scarce persistent monitoring.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Why 3D Reconstruction Breaks Down on Under-Ice AUV Video, and How to Predict It<\/a><\/h4>\n\n\n\n

Ana-\u0218tefania Kog\u0103lniceanu, Klara Orban, Ashwin Nedungadi, Stefan L\u00fcdtke, Kuderna-Iulian Ben\u021ba<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Reconstruction of 3D structures using under-ice AUV video footage has been inaccurate, although it is not certain whether this is due to shortcomings in the methods or an inherent deficiency of the capture system. We test four different models for reconstruction on the same Antarctic under-ice AUV video (Icefin, Kamb Ice Stream): classical feature matching SfM, differentiable photometric splatting, volumetric depth fusion, and a pose-free feed-forward foundation model. All four fail, each in a distinct way: near-total registration yielding almost no structure, geometrically broken output despite healthy photometric metrics, a persistent coverage artifact, and unstable cross-chunk scale. We show that all four are consistent with the same capture regime, a small camera-baseline-to-scene-depth ratio. Following our analysis, we propose a detection system that flags this type of failure in minutes, rather than hours of wasted compute.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

A Practical Framework for Building Benthic Segmentation Models from Few Annotated Images<\/a><\/h4>\n\n\n\n

Mara Sangiovanni, Sabrina Rossi<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Pixel-level annotation of benthic imagery is labor-intensive and often limits the application of deep learning to local marine monitoring campaigns. In this work, we investigate a practical training pipeline for semantic segmentation of Mediterranean benthic communities under an extreme low-data regime, where only a few annotated images are available alongside a larger collection of unlabeled data. Rather than proposing a new architecture, we systematically evaluate transfer learning strategies, class-imbalance handling, target-domain self-supervised pretraining through masked image modeling, and iterative pseudo-labeling. Our experiments show that generic marine pretraining is not necessarily transferable to a specific ecological domain and that the effectiveness of training strategies strongly depends on the interaction between architecture, supervision, and dataset preparation. The proposed workflow achieves a final mIoU of approximately 0.48 despite relying on fewer than 50 annotated images, demonstrating that practical segmentation models can be developed even under severe annotation constraints.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Multimodal Taxonomic Conditioning for Generative Plankton Imagery<\/a><\/h4>\n\n\n\n

Daniela Ivanova, Ozgu Goksu, Nicolas Pugeault<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Automated plankton imaging produces severely long-tailed datasets, where the rare taxa of greatest ecological interest have too few images to train or evaluate classifiers reliably. We generate synthetic plankton imagery conditioned on taxonomy: a CLIP encoder is adapted on a large plankton corpus with a ranked contrastive objective extended to deep, ragged taxonomies, then frozen to condition a parameter-efficient diffusion transformer. We evaluate synthetic sample quality on distributional fidelity and downstream classifier utility.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

SCUBA: Beyond Per-Frame Metrics for Closed-Loop Underwater Autonomy<\/a><\/h4>\n\n\n\n

Shayok Bagchi, Mohamed Saad Ibnseddik, Abrar Majeedi, Yin Li<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Underwater perception degrades with turbidity, wavelength-dependent attenuation, and low light, yet the robustness of the resulting autonomy is usually evaluated open loop: corruptions are applied to static images and per-frame task performance metrics are reported. However, autonomous underwater operations demand closed-loop operation between perception and control elements. We answer a key open question in this paper: do per-frame perception metrics predict mission outcomes in closed-loop perception-control autonomous missions? We present a systematic comparison between open-loop proxy metrics and closed-loop mission performance in underwater visual navigation. We collect over 1,500 annotated real pool images from a BlueROV2-class vehicle, and release our dataset, degradation suite, and evaluation harness. Across 16 conditions and 480 closed-loop trials, we find that per-frame detection rate is a poor predictor of mission outcome: an apparent correlation of r = 0.78 collapses to 0.27 once a single catastrophic condition is excluded, rank correlation is statistically indistinguishable from zero, and the errors run in both directions, with one condition sustaining 30\/30 mission success at a 38.6% detection rate while another fails 12\/30 at 64.0%. A serendipitous hardware comparison further suggests that camera capture rate materially affects mission success.<\/p>\n<\/details>\n<\/div>\n<\/div>\n\n\n\n

\n
\n
\"\"\n\t\t\t\n\t\t\t\t\n\t\t\t<\/svg>\n\t\t<\/button><\/figure>\n<\/div>\n\n\n\n
\n

Quantum-Gated LiteSSD: A Parameter-Efficient Lightweight Hybrid Quantum-Classical Framework for Forward-Looking Sonar Object Detection<\/a><\/h4>\n\n\n\n

Niloy Kumar Mondal, Poulomi Sarker Puja<\/em><\/p>\n\n\n\n

Abstract<\/summary>\n

Forward-looking sonar object detection is essential for underwater perception, yet deployment on embedded platforms requires highly compact models. To address this challenge, we explore quantum computing and introduce Quantum-Gated LiteSSD<\/strong>, a parameter-efficient hybrid quantum-classical detector that reformulates QuCNet-style multi-circuit quantum processing as an identity-centred channel-gating mechanism for spatial feature modulation. Experiments on the Marine Debris Watertank dataset and UATD forward-looking sonar benchmarks demonstrate an effective parameter-accuracy trade-off. The proposed detector achieves 90.84% on Watertank with approximately 62\u00d7 fewer parameters than YOLO26s and 164.3\u00d7 fewer than SSD-VGG16. On UATD, the model achieves 70.37% with only 0.150M parameters, making it approximately 4.1\u00d7 smaller than SSGA-YOLO while retaining meaningful multi-class detection capability.<\/p>\n<\/details>\n<\/div>\n<\/div>\n","protected":false},"excerpt":{"rendered":"

This is the list of short papers accepted for presentation at the 2nd Workshop on Marine Vision 2026. The papers are publicly available through OpenReview. The papers appear in random order. Semantic Late Interaction for Cross-Temporal Reef Relocalization Hugues Sibille, Jonathan Sauder, Guilhem Banc-Prandi, Devis Tuia Marine Wildlife Individual Re-identification Using Deep Learning: A Component […]<\/p>\n","protected":false},"author":3,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-627","page","type-page","status-publish","hentry"],"_links":{"self":[{"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/pages\/627","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/users\/3"}],"replies":[{"embeddable":true,"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/comments?post=627"}],"version-history":[{"count":6,"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/pages\/627\/revisions"}],"predecessor-version":[{"id":672,"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/pages\/627\/revisions\/672"}],"wp:attachment":[{"href":"https:\/\/vap.aau.dk\/marinevision\/wp-json\/wp\/v2\/media?parent=627"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}