Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark

Martin Kvisvik Larsen; Oscar Pizarro

cs.CVMar 4, 2026

Long-Term Visual Localization in Dynamic Benthic Environments: A Dataset, Footprint-Based Ground Truth, and Visual Place Recognition Benchmark

Authors: Martin Kvisvik Larsen, Oscar Pizarro

Organizations: Department of Marine Technology, Norwegian University of Science and Technology, Trondheim, Norway

Abstract

Long-term visual localization has the potential to reduce cost and improve mapping quality in optical benthic monitoring with autonomous underwater vehicles (AUVs). Despite this potential, long-term visual localization in benthic environments remains understudied, primarily due to the lack of curated datasets for benchmarking. Moreover, limited georeferencing accuracy and image footprints necessitate precise geometric information for accurate ground-truthing. In this work, we address these gaps by presenting a curated dataset for long-term visual localization in benthic environments and a novel method to ground-truth visual localization results for near-nadir underwater imagery. Our dataset comprises georeferenced AUV imagery from five benthic reference sites, revisited over periods up to six years, and includes raw and color-corrected stereo imagery, camera calibrations, and sub-decimeter registered camera poses. To our knowledge, this is the first curated underwater dataset for long-term visual localization spanning multiple sites and photic-zone habitats. Our ground-truthing method estimates 3D seafloor image footprints and links camera views with overlapping footprints, ensuring that ground-truth links reflect shared visual content. Building on this dataset and ground truth, we benchmark eight state-of-the-art visual place recognition (VPR) methods and find that Recall@K is significantly lower on our dataset than on established terrestrial and underwater benchmarks. Finally, we compare our footprint-based ground truth to a traditional location-based ground truth and show that distance-threshold ground-truthing can overestimate VPR Recall@K at sites with rugged terrain and altitude variations. Together, the curated dataset, ground-truthing method, and VPR benchmark provide a stepping stone for advancing long-term visual localization in dynamic benthic environments.

Figures & tables

Site	Geohash	Visit Years	Depth range	Image Count	Site Description
-	-	-	m	-	-
Site 1	QDCH0FTQ	2010, 2011, 2012, 2013	17 - 20	17794	Dense coral reef
Site 2	QDCHDMY1	2011, 2012, 2013, 2017	27 - 35	20944	Dense coral reef and soft sediment
Site 3	R23685BC	2010, 2012, 2014	40 - 43	15146	Rock reef with corals and sponges and sandy substrate
Site 4	R29MRD5H	2009, 2011, 2013	25 - 36	21362	Boulder reef
Site 5	R7JJSKXQ	2010, 2012, 2013	18 - 21	19444	Rock reef with corals and sponges

Table 1: Overview of key properties for the benthic reference sites in our dataset.

Figure 1: Map with overview of the benthic reference sites in our dataset. (A) Map of Australia and the geographic location of each site. (B-F) Orthomosaics showing seafloor patches with characteristics representative for the following benthic reference sites; (B) Site 1, (C) Site 2, (D) Site 3, (E) Site 4, and (F) Site 5. The orthomosaic patches are rendered at the same spatial scale, indicated by the scalebar in the lower right corner. Base map data © Commonwealth of Australia (Australian Bureau of Statistics) 2021, ASGS Edition 3 digital boundaries, used under CC BY 4.0 ( Australian Bureau of Statistics, 2021 ) .

Figure 2: (A) Raw (left) and color corrected (right) image from the 2010 visit to Site 1. (B) Raw (left) and color corrected (right) image from the 2013 visit to Site 1.

Figure 3: Single line diagram of the geometric reconstruction and registration workflow for an example with two visits. In the example Visit 1 is the registration target, while Visit 2 is the registration source being registered to Visit 1. Note that for the target visit, the SFM -estimated camera calibrations and interpolated camera poses are the final calibrations and camera poses, while for the source visit they are corrected by the similarity transformation estimated by the geometric registration.

Figure 4: An example showing the fusion of relative and metric stereo-derived range maps for an image from 2009 visit to Site 4. (A) The color corrected image, (B) the stereo range map reprojected into the pixel coordinate system of the original image, and (C) the range map estimated by fusing the relative range map with the stereo-derived range map. The stereo-derived and fused range maps are rendered with the same color map to indicate the range. Overall, the fused range maps show sharper edges and do not exhibit the geometric inconsistencies on the left side seen in the stereo range map, which are believed to be caused by limited left–right image overlap.

Figure 5: Illustration of a 2D simplified model of our footprint estimation method for two scenarios. Panel (A) shows a scenario where local terrain relief causes non-overlapping image footprints for two spatially close camera views. Panel (B) shows a scenario where large altitude differences causes spatially distant camera views to have overlapping image footprints.

Figure 6: Distribution of geometric registration errors across the benthic reference sites in the dataset. The legends are on the format "Target-Source", and indicate the year of the target and source visit for each visit pair used in the geometric registration.

Site	Visit Years	Database View Count	Query View Count	Avg. Database View Alt.	Avg. Query View Alt.	Query Coverage Overlap	Link Count	ALQ *	Avg. Footprint Width	Loc. Dist. 95th Perc.
-	-	-	-	m	m	-	-	-	m	m
Site 1	2010 – 2011	2323	2255	2.04	2.04	74 %	27.8 K	15	1.65	1.37
	2010 – 2012	2323	2184	2.04	2.00	53 %	14.8 K	13	1.65	1.37
	2010 – 2013	2323	2134	2.04	1.99	65 %	21.1 K	15	1.64	1.37
	2011 – 2012	2255	2184	2.04	2.00	65 %	16.4 K	10	1.64	1.41
	2011 – 2013	2255	2134	2.04	1.99	80 %	21.0 K	11	1.63	1.36

Table 2: Overview of the visit pairs and footprint links filtered with

\tau_{f}=0.07

. The visit pairs are created using the temporal pairing strategy outlined in Section 2.5.2 . * Average link per valid query.

Figure 7: Mean Recall@K across all visit pairs for each site in the dataset. Results are based on the footprint-based ground truth from Section 3.2 .

Methods	Site 1		Site 2		Site 3		Site 4		Site 5
Methods	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10
	%	%	%	%	%	%	%	%	%	%
AnyLoc ( Keetha et al., 2024 )	24.9	51.2	7.4	28.2	10.0	30.8	21.4	49.9	10.1	31.9
CliqueMining ( Izquierdo and Civera, 2025 )	17.0	39.8	5.6	24.6	10.9	32.3	13.5	32.4	7.2	25.6
CosPlace ( Zaccone et al., 2024 )	13.8	37.1	4.3	20.1	5.5	22.9	4.9	19.1	5.7	22.3
EigenPlaces ( Berton et al., 2023 )	14.7	38.0	4.9	22.8	6.0	22.7	6.0	23.5	5.9	23.8

Table 3: Recall@1 and Recall@10 averaged across every visit pair for the reference sites in the dataset. For each rank cutoff

K

and site, we use green and red to highlight the best and second best model, respectively. Results are based on the footprint-based ground truth from Section 3.2 .

Figure 8: Visual place recognition results at Site 1 rendered on top of an orthomosaic derived from the 2010 images, using MegaLoc with $K=5$ to retrieve 2010 database images for 2013 query images. (A) Overview of place recognition results for the entire site. (B) Cluster of recognized query camera views at the end of a survey leg. (C) Cluster of recognized query camera views at a cross-over point between survey legs. (D) Trajectory segment with multiple consecutively recognized query camera views.

Figure 9: Visual place recognition results at Site 2 rendered on top of an orthomosaic derived from the 2011 images, using MegaLoc with $K=5$ to retrieve 2011 database images for 2013 query images. Panel (A) show an overview of place recognition results for the entire site. Panel (B) shows a cluster of recognized query camera views at the end of a survey leg. Panel (C) shows trajectory segments with recognized query camera views as the AUV passes over coral colonies in the transition between the soft sediment bottom and the dense coral reef. Panel (D) shows trajectory segments with recognized query camera views as the AUV passes over a seagrass meadow on the soft sediment bottom.

Figure 10: Recall@10 for all VPR models for varying revisit intervals to the sites in the dataset. The overall trend shows a decline of VPR Recall@10 with increasing revisit intervals. For Site 1 and Site 2, Recall@10 shows a sharp decline between revisit intervals of 1 and 2 years, before a more gradual decline. For Site 5, Recall@10 increases between revisit intervals of 2 and 3 years, indicating that additional factors affect Recall@10. Results are based on the footprint-based ground truth from Section 3.2 .

Figure 11: Scatter plots of (A) R@10 and (B) IR@10 for the footprint-based and location-based ground truth. The R@10 is consistently higher for the location-based ground truth for all sites and VPR models, and indicates that model performance is over-estimated with the location-based ground truth. IR@10 is consistently higher for the footprint-based ground truth for all sites and models, and indicate that the VPR retrievals are more aligned with the footprint-based links.

Site	Visit Years	Foot. Width	Dist. Thres.	Foot. ALQ	Loc. ALQ	MegaLoc Foot. Links		MegaLoc Loc. Links		NetVLAD Foot. Links		NetVLAD Loc. Links
Site	Visit Years	Foot. Width	Dist. Thres.	Foot. ALQ	Loc. ALQ	IR@10	R@10	IR@10	R@10	IR@10	R@10	IR@10	R@10
-	-	m	m	-	-	%	%	%	%	%	%	%	%
Site 1	2010 – 2011	1.65	1.37	15	18	9.8	59.5	8.7	60.5	3.1	29.1	2.9	31.2
	2010 – 2012	1.65	1.37	13	15	6.8	47.2	6.1	49.1	1.7	18.3	1.6	19.4
	2010 – 2013	1.64	1.37	15	17	5.0	36.3	4.5	37.8	1.8	20.9	1.8	22.7
	2011 – 2012	1.64	1.41	10	13	9.6	46.1	8.1	46.7	2.6	16.8	2.6	19.2

Table 4: Summary of geometric statistics and retrieval performance for MegaLoc and NetVLAD under footprint-based and location-based ground truth. Footprint-based ground-truth links from Section 3.2 . For each visit pair, the location-based distance threshold is set to the 95th percentile of distances between footprint-linked camera views. “Foot. Width” and “Dist. Thres.” give the average footprint width and the resulting distance threshold, respectively. “Foot. ALQ” and “Loc. ALQ” report the average number of links per valid query under the two ground truths. R@10 and IR@10 denote

\mathrm{Recall@K}

and

\mathrm{IRRecall@K}

K=10

, respectively.

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Registration Stage	Estimation Method	Parameters	Configuration 1	Configuration 2
Stage 1	FPFH-RANSAC	Voxel Size	0.08 m	0.20 m
Stage 1	FPFH-RANSAC	Distance Threshold	0.06 m	0.15 m
Stage 2	Colored ICP	Voxel Size	0.10 m	0.20 m
Stage 2	Colored ICP	Distance Threshold	0.15 m	0.20 m
Stage 3	Colored ICP	Voxel Size	0.05 m	0.05 m
Stage 3	Colored ICP	Distance Threshold	0.05 m	0.05 m

Appendix

Table 5: The two configurations of the geometric registration pipeline that is used to register the visits for each site in our dataset. Configuration 1 is used to register the visits for Site 1, while Configuration 2 is used to register the visits to Site 2, 3, 4, and 5. Overall, Configuration 1 performs registration at finer spatial resolutions than Configuration 2.

Site	Visit Years	AnyLoc		CliqueMining		CosPlace		EigenPlaces		MegaLoc		MixVPR		NetVLAD		SALAD
Site	Visit Years	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10	R@1	R@10
Site 1	2010-2011	42.3	73.0	30.3	56.2	20.1	49.1	21.0	49.3	31.3	59.5	12.6	35.7	5.3	29.1	30.4	54.8
	2010-2012	20.9	50.9	14.4	40.0	14.0	36.2	13.2	37.4	19.2	47.2	6.2	22.1	2.9	18.3	15.2	41.2
	2010-2013	15.2	42.3	10.8	34.1	10.3	33.0	12.4	35.6	13.1	36.3	7.4	28.8	4.5	20.8	13.7	37.2
	2011-2012	24.2	47.6	16.7	39.0	11.6	32.9	12.6	33.4	23.0	46.1	6.6	21.3	3.6	16.8	19.4	40.8
	2011-2013	14.2	35.7	8.3	27.6	10.2	27.7	10.9	29.2	13.8	33.9	4.7	19.7	4.8	18.5	10.0	28.8

Appendix

Table 6: Recall@1 and Recall@10 for all VPR models and visit pairs in the dataset. For each rank cutoff

K

and visit pair, we use green and red to highlight the best and second best model, respectively.

Figure 12: Visual place recognition results at Site 3 rendered on top of an orthomosaic derived from the 2010 images, using MegaLoc with $K=5$ to retrieve 2010 database images for 2012 query images.

Figure 13: Visual place recognition results at Site 4 rendered on top of an orthomosaic derived from the 2009 images, using MegaLoc with $K=5$ to retrieve 2009 database images for 2011 query images.

Figure 14: Visual place recognition results at Site 5 rendered on top of an orthomosaic derived from the 2010 images, using MegaLoc with $K=5$ to retrieve 2010 database images for 2013 query images.

Figure 15: Examples of very hard, hard, and easy query images from each benthic reference site. The very hard queries are not recognized by any VPR model, hard queries are recognized by the best models, i.e. MegaLoc and AnyLoc, and easy queries are recognized by all models.

Figure 16: Examples illustrating disagreement between footprint-based and location-based linking of query–database camera views. Location-based linking used a threshold corresponding to the 95th percentile of spatial distances between footprint-linked camera views. Panels (A) and (B) show image pairs that are linked by overlapping footprints but not by location proximity. These cases involve large altitude differences between the query and database camera views, preventing location-based linking. Manually added markers highlight visual features shared between the paired images. Panels (C) and (D) show image pairs linked by location proximity but not by overlapping footprints. Here, abrupt changes in seafloor relief cause the cameras to operate at low altitudes, limiting the field of view. Although these image pairs are spatially close, they lack shared visual features.

Robot Localization

Robot localization estimates a robot’s position and orientation from onboard sensors, maps, and observations of its surroundings.

See all 175 papers

Oct 6, 2025

BenthiCat: An opti-acoustic dataset for advancing benthic classification and habitat mapping

Benthic habitat mapping is fundamental for understanding marine ecosystems, guiding conservation efforts, and supporting sustainable resource management. Yet, the scarcity of large, annotated datasets limits the development and benchmarking of machine learning models in this domain. This paper introduces a thorough multi-modal dataset, comprising about a million side-scan sonar (SSS) tiles collected along the coast of Catalonia (Spain), complemented by bathymetric maps and a set of co-registered optical images from targeted surveys using an autonomous underwater vehicle (AUV). Approximately 36000 of the SSS tiles have been manually annotated with segmentation masks to enable supervised fine-tuning of classification models. All the raw sensor data, together with mosaics, are also released to support further exploration and algorithm development. To address challenges in multi-sensor data fusion for AUVs, we spatially associate optical images with corresponding SSS tiles, facilitating self-supervised, cross-modal representation learning. Accompanying open-source preprocessing and annotation tools are provided to enhance accessibility and encourage research. This resource aims to establish a standardized benchmark for underwater habitat mapping, promoting advancements in autonomous seafloor classification and multi-sensor integration.

Hayat Rajani, Valerio Franchi, Borja Martinez-Clavel Valles +3

Computer Vision and Robotics Research Institute, University of Girona, Spain · Tecnoambiente SLU, Spain

Underwater Imaging

Ask PDF

Jun 23, 2026

From Open Waters to Enclosed Cabins: ProteusVPR for Cross-Scene Visual Place Recognition in Maritime Perception and Cabin Inspection

Autonomous robotic inspection in maritime environments presents unique challenges for Visual Place Recognition (VPR) due to cross-scene perceptual shifts. Robots navigating ship-borne environments must transition between visually distinct domains: open decks with sparse textures and severe illumination changes, and enclosed cabins with repetitive structures and high visual ambiguity. Existing VPR methods, designed primarily for urban or indoor scenes, fail to generalize reliably across these starkly different scenarios. To address this, we propose ProteusVPR, a two-stage retrieval-refinement framework. The first stage employs any standard VPR model for initial image retrieval. The second stage introduces a geometric-visual estimation network that fuses the retrieved image with two temporally preceding frames, incorporating geometric descriptors, a local affine coordinate system, and camera azimuth encoding to achieve precise localization. To support this task, we introduce the XHZ dataset, an 8K-panoramic ship-borne dataset collected from an operational vessel, featuring multi-floor cabin structures, deck transition zones, and strict query-database separation for rigorous evaluation. Extensive experiments on the XHZ dataset demonstrate that ProteusVPR consistently improves the localization accuracy across multiple VPR backbones, reducing mean localization error by over 60% on average and that ProteusVPR offers an effective and robust solution for precise visual localization in challenging, cross-scene maritime environments.

Zexi Chen, Zitai Huang, Qiwen Gu +6

Shanghai Jiaotong University, Shanghai, China · Shanghai Ship and Shipping Research Institute Co.LTD, Shanghai, China · Tongji University, Shanghai, China +2

Visual Place Recognition Robotics Robot Localization

Ask PDF

Jul 6, 2026

DIVO: Continuous-time DVL-Inertial-Visual Odometry for Unmanned Underwater Vehicles

This paper presents a novel acoustic-visual-inertial odometry solution leveraging a continuous-time trajectory estimation framework for unmanned underwater vehicles. Underwater environments present unique challenges for visual localization and mapping, such as light attenuation, illumination variance, and the presence of particulate matter. This motivates the use of additional sensing modalities and a visual tracking pipeline that is robust to diverse subsea conditions. The proposed system is the first continuous-time trajectory estimation framework based on Gaussian processes to fuse asynchronous measurements from a Doppler velocity log, a stereo camera, and an inertial measurement unit. Additionally, a novel visual frontend is proposed, incorporating learning-based feature extraction and matching that is robust to the specific challenges that subsea environments present. The proposed framework enables seamless integration of additional sensor modalities in continuous-time and is adaptable to different environments without reconfiguration. The proposed system is extensively tested on real-world underwater inspection datasets, where it outperforms state-of-the-art visual-inertial and acoustic-visual-inertial SLAM algorithms in accuracy, robustness, and trajectory coverage. Notably, the proposed system outperforms the state-of-the-art despite only forming short-term visual data associations.

Kyungmin Jung, Angad Bajwa, Junha Yoo +2

Department of Mechanical Engineering, McGill University, Montreal, QC, Canada

Underwater Robotics Visual-Inertial Odometry Autonomous Underwater Navigation Robot State Estimation Multimodal Sensor Fusion

Ask PDF

Abstract

Robot Localization

Explore similar work

BenthiCat: An opti-acoustic dataset for advancing benthic classification and habitat mapping

From Open Waters to Enclosed Cabins: ProteusVPR for Cross-Scene Visual Place Recognition in Maritime Perception and Cabin Inspection

DIVO: Continuous-time DVL-Inertial-Visual Odometry for Unmanned Underwater Vehicles