Open-set multi-view seahorse re-identification from underwater photograph sequences

Tristan Fauvel; Sylvie Louisy; Patrick Louisy; Thomas Fauvel

cs.CVOct 29, 2026

Open-set multi-view seahorse re-identification from underwater photograph sequences

Authors: Tristan Fauvel, Sylvie Louisy, Patrick Louisy, Thomas Fauvel

Organizations: Baysic Labs 37 rue Sommeiller, 74000 Annecy, France · Association Peau-Bleue 46 Rue des Escais, 34300 Agde, France · UMR BOREA (MNHN, CNRS, Sorbonne Université, IRD, UniCaen, UGA), Station marine de Dinard, CRESCO 38 rue du Port Blanc, 35800 Dinard, France · Parc Naturel Marin du Bassin d’Arcachon, 4 Rue Copernic, 33470 Le Teich, France ·

Abstract

Automated animal re-identification from photograph sequences can facilitate minimally invasive Capture-Mark-Recapture studies at scale. Yet, photographs taken during an observation session can vary greatly in quality, not always providing the necessary information for accurate re-identification. In this paper, we present new methods for taxonomic classification, sex determination, and open-set re-identification from image sequences. We train and evaluate our models on a dataset of 35,914 photographs of 474 individual seahorses observed in the wild. We demonstrate that our approach, by combining information from multiple photographs, yields state-of-the-art performance.

Figures & tables

Figure 1: Illustrative examples of photographs from the dataset. As this dataset was not specifically collected for the purpose of automated re-identification, images vary greatly in quality. Some images focus on the seahorse’s head, while others show the individual in its environment, with potential occlusion. Some images were taken for the purpose of measuring the animal’s length.

Species	Sex	N	FU med/max (d)	Obs/ind. (IQR)	Photos/obs (IQR)
H. guttulatus	Male	213	40/1015	3 (1–7)	13 (10–17)
H. guttulatus	Female	185	40/1094	2 (1–8)	13 (10–16)
H. guttulatus	Unknown	63	0/41	1 (1–1)	15 (12–20)
H. hippocampus	Male	8	2/87	2 (1–2)	16 (13–41)
H. hippocampus	Female	5	39/108	5 (5–7)	13 (10–32)

Table 1: Summary statistics by species and sex. FU: follow-up time in days. Obs/ind.: number of observations per individual. Photos/obs: number of photos per observation. IQR: interquartile range.

Species	n gaps	Min (d)	P25 (d)	Median (d)	P75 (d)	Max (d)	Mean $\pm$ SD
H. guttulatus	1983	1	7	13	31	609	33 $\pm$ 58
H. hippocampus	36	1	1	7	14	77	12 $\pm$ 17

Table 2: Inter-observation gap statistics by species. SD: standard deviation. P25: 25th percentile. P75: 75th percentile.

Method	Balanced acc. $\uparrow$	Macro F1 $\uparrow$
DINOv3 ViT-L/16 CLS (image)	0.88 [0.84, 0.91]	0.87 [0.81, 0.90]
DINOv3 ViT-L/16 CLS (observation)	0.94 [0.85, 1.00]	0.97 [0.91, 1.00]
BioCLIP 2.5 (image)	0.82 [0.75, 0.88]	0.74 [0.66, 0.78]
BioCLIP 2.5 (observation)	0.90 [0.80, 0.98]	0.94 [0.87, 0.98]

Table 3: Backbone comparison at the image and observation levels (observation = mean-logit fusion) for species classification. All metrics are identity-balanced, with 95% individual-level bootstrap intervals over the 13 H. hippocampus and 461 H. guttulatus individuals.

Method	Balanced acc. $\uparrow$	Macro F1 $\uparrow$	Brier $\downarrow$	ECE $\downarrow$	MCE $\downarrow$
Image (LR)	0.88 [0.84, 0.91]	0.87 [0.81, 0.90]	0.01 [0.01, 0.02]	0.01 [0.01, 0.01]	0.42 [0.29, 0.69]
Mean logit	0.94 [0.85, 1.00]	0.97 [0.91, 1.00]	1.30 [0.02, 3.63] ${}\times 10^{-3}$	2.20 [0.21, 6.17] ${}\times 10^{-3}$	0.58 [0.22, 0.59]
Logit sum	0.94 [0.85, 1.00]	0.97 [0.91, 1.00]	2.93 [0.00, 8.80] ${}\times 10^{-3}$	3.04 [0.00, 9.12] ${}\times 10^{-3}$	0.75 [0.00, 0.75]
Majority vote	0.99 [0.97, 1.00]	1.00 [0.99, 1.00]	3.12 [1.74, 5.10] ${}\times 10^{-3}$	0.01 [0.01, 0.02]	0.60 [0.44, 0.60]

Table 4: Comparison of image fusion methods with DINOv3 CLS, with the per-image logistic regression as a baseline, in species classification. All metrics are identity-balanced, with 95% individual-level bootstrap intervals. ECE: Expected Calibration Error. MCE: Maximum Calibration Error.

Method	Balanced acc. $\uparrow$	Macro F1 $\uparrow$	Brier $\downarrow$	ECE $\downarrow$	MCE $\downarrow$
Image (LR)	0.79 [0.77, 0.81]	0.79 [0.77, 0.80]	0.15 [0.14, 0.16]	0.06 [0.04, 0.07]	0.11 [0.08, 0.17]
Mean logit	0.89 [0.87, 0.92]	0.89 [0.87, 0.92]	0.08 [0.07, 0.10]	0.07 [0.05, 0.09]	0.13 [0.12, 0.30]
Logit sum	0.89 [0.87, 0.92]	0.89 [0.87, 0.92]	0.10 [0.08, 0.12]	0.10 [0.08, 0.12]	0.67 [0.50, 0.79]
Majority vote	0.87 [0.85, 0.90]	0.87 [0.84, 0.90]	0.10 [0.08, 0.11]	0.05 [0.04, 0.08]	0.16 [0.13, 0.27]

Table 5: Comparison of image fusion methods with DINOv3 CLS, with the per-image logistic regression as a baseline, in sex classification. All metrics are identity-balanced, with 95% individual-level bootstrap intervals. ECE: Expected Calibration Error. MCE: Maximum Calibration Error.

Method	BAKS $\uparrow$	BAUS $\uparrow$	BRA $\uparrow$	NA $\uparrow$
1. MiewID zero-shot	0.22 [0.16, 0.28] (-)	0.62 [0.57, 0.68] (-)	0.47 [0.42, 0.53] (-)	0.33 [0.26, 0.38] (-)
2. + MiewID fine-tuning	0.61 [0.54, 0.68] (+0.40)	0.65 [0.59, 0.71] (+0.02)	0.64 [0.59, 0.68] (+0.16)	0.61 [0.56, 0.66] (+0.29)
3. + relevance weighting	0.59 [0.52, 0.65] (-0.02)	0.66 [0.61, 0.71] (+0.01)	0.64 [0.59, 0.68] (-0.00)	0.60 [0.55, 0.65] (-0.01)
4. + ALIKED/LightGlue rerank	0.62 [0.53, 0.70] (+0.03)	0.86 [0.81, 0.91] (+0.20)	0.77 [0.73, 0.81] (+0.14)	0.72 [0.65, 0.77] (+0.12)
5. + logistic knownness	0.67 [0.60, 0.74] (+0.05)	0.78 [0.74, 0.82] (-0.08)	0.74 [0.69, 0.78] (-0.04)	0.71 [0.65, 0.75] (-0.01)
6. + MAP open-set decision	0.63 [0.56, 0.70] (-0.04)	0.87 [0.83, 0.90] (+0.09)	0.78 [0.73, 0.82] (+0.04)	0.72 [0.66, 0.76] (+0.01)

Table 6: Seahorse re-ID with a sparse gallery. Cells are means over 5 query/gallery role rotations within the test split, with 95% bootstrap intervals over test individuals in brackets and the change from the row above in parentheses.

Method	BAKS $\uparrow$	BAUS $\uparrow$	BRA $\uparrow$	NA $\uparrow$
1. MiewID zero-shot	0.34 [0.25, 0.42] (-)	0.58 [0.50, 0.65] (-)	0.49 [0.42, 0.56] (-)	0.40 [0.32, 0.47] (-)
2. + MiewID fine-tuning	0.68 [0.58, 0.77] (+0.34)	0.71 [0.64, 0.78] (+0.13)	0.70 [0.64, 0.76] (+0.20)	0.67 [0.60, 0.73] (+0.27)
3. + relevance weighting	0.70 [0.61, 0.79] (+0.02)	0.86 [0.81, 0.91] (+0.15)	0.80 [0.76, 0.84] (+0.11)	0.77 [0.71, 0.81] (+0.09)
4. + ALIKED/LightGlue rerank	0.78 [0.69, 0.86] (+0.08)	0.71 [0.65, 0.77] (-0.15)	0.74 [0.68, 0.79] (-0.07)	0.71 [0.64, 0.76] (-0.05)
5. + logistic knownness	0.84 [0.75, 0.91] (+0.06)	0.93 [0.89, 0.97] (+0.22)	0.90 [0.85, 0.93] (+0.16)	0.88 [0.82, 0.92] (+0.17)
6. + MAP open-set decision	0.82 [0.73, 0.89] (-0.02)	0.94 [0.90, 0.97] (+0.01)	0.90 [0.85, 0.93] (-0.00)	0.87 [0.82, 0.92] (-0.01)

Table 7: Seahorse re-ID with a rich gallery. Cells are means over 5 query/gallery role rotations within the test split, with 95% bootstrap intervals over test individuals in brackets and the change from the row above in parentheses.

Figure 2: Example images of high relevance (upper row) versus low-relevance (lower row) images.

Figure 3: True positive rate as a function of the false positive rate for determining whether the query corresponds to a new individual or an individual in the gallery, for the maximum-score thresholding method (orange curve) or for the logistic knownness detection model (blue curve). The left plot corresponds to the sparse gallery setting, whereas the right plot corresponds to the rich gallery case.

Animal Re-Identification

Animal re-identification matches observations of the same individual across time, cameras, viewpoints, and changing conditions.

See all 34 papers

Calibrated Similarity and Graph Clustering for Open-Set Animal Re-Identification
Aug 3, 2026Mohamed ElBassat, Seifeldin Elkerdany, Mohamed ElBialy +5Animal Re-Identification Image Matching
Made In Alexandria Artificial Intelligence Team, Alexandria, Egypt · Faculty of Computers and Data Science, Alexandria University, Alexandria, Egypt · Faculty of Computer Science and Engineering, Alamein International University, New Alamein City, 51718, Egypt +3
Patch Ensembles for Robust Salmon Re-Identification with Weak Trajectory Labels
May 18, 2026Espen Uri Høgstedt, Christian Schellewald, Annette Stahl +1Biometric Identification Animal Re-Identification
Department of Computer Science, Norwegian University of Science and Technology, Norway · SINTEF Ocean, Norway · Department of Engineering Cybernetics, Norwegian University of Science and Technology, Norway
DS@GT ARC at AnimalCLEF 2026: Species-Aware Graph Construction for Multi-Species Animal Re-Identification
Jul 17, 2026Evan Sinclair Smith, Anthony Miyaguchi, Snigdha Palamari +1Animal Re-Identification Image Matching
1Georgia Institute of Technology, North Ave NW, Atlanta, GA 30332

Aug 3, 2026J/K move · Enter open · S save

Calibrated Similarity and Graph Clustering for Open-Set Animal Re-Identification

Mohamed ElBassat, Seifeldin Elkerdany, Mohamed ElBialy, Gamal Abouelhamd+4

Made In Alexandria Artificial Intelligence Team, Alexandria, Egypt · Faculty of Computers and Data Science, Alexandria University, Alexandria, Egypt · Faculty of Computer Science and Engineering, Alamein International University, New Alamein City, 51718, Egypt+3

Animal Re-Identification Image Matching Open-Set Recognition Image Clustering

Open paper AskPDF arXiv

AnimalCLEF26 addresses discovery-oriented animal re-identification, where systems must both attach query images to known individuals and discover unseen individuals by clustering them correctly. We present a similarity-to-clustering pipeline for this setting across Eurasian lynx, fire salamander, loggerhead sea turtle, and Texas horned lizard images. The method first isolates the target specimen using segmentation and then applies lightweight species-specific preprocessing for lynx, sea turtle, and salamander images to enhance identity-relevant visual cues, while Texas horned lizard images are used after segmentation only. Pairwise similarities are then estimated with WildFusion by calibrating and combining a MiewID global descriptor with two local matching branches, ALIKED + LightGlue and DISK + LightGlue. The resulting query-query similarities are refined and converted into identity clusters using graph-based clustering, while query-database similarities are used to attach confident samples to known identities. We evaluate training-free and fine-tuned MiewID variants, including Dynamic ArcFace and SphereFace2-Focal adaptations, and combine them in the final ensemble. Our selected ensemble substantially improves on the WildFusion baseline, achieving the best public ARI of 0.72124 and a private ARI of 0.70393, while a simpler preprocessing-before-calibration variant achieves the best private ARI of 0.71087. These results indicate that calibrated global-local fusion with species-aware preprocessing choices is effective for open-set wildlife re-identification under challenging field conditions and visual variation. The implementation code is available on GitHub.

Abstract

Animal Re-Identification

Explore similar work

Calibrated Similarity and Graph Clustering for Open-Set Animal Re-Identification

Patch Ensembles for Robust Salmon Re-Identification with Weak Trajectory Labels

DS@GT ARC at AnimalCLEF 2026: Species-Aware Graph Construction for Multi-Species Animal Re-Identification