From Visual to Multimodal: Systematic Ablation of Encoders and Fusion Strategies in Animal Identification

Vasiliy Kudryavtsev; Kirill Borodin; German Berezin; Kirill Bubenchikov; Grach Mkrtchian; Alexander Ryzhkov

cs.CVFeb 28, 2026

From Visual to Multimodal: Systematic Ablation of Encoders and Fusion Strategies in Animal Identification

Authors: Vasiliy Kudryavtsev, Kirill Borodin, German Berezin, Kirill Bubenchikov, Grach Mkrtchian, Alexander Ryzhkov

Organizations: Faculty of IT, Technical University of Communication and Informatics, Moscow, Russia, 111024 · AI lab, Avito, Moscow, Russia, 125196

Abstract

Automated animal identification is a practical task for reuniting lost pets with their owners, yet current systems often struggle due to limited dataset scale and reliance on unimodal visual cues. This study introduces a multimodal verification framework that enhances visual features with semantic identity priors derived from synthetic textual descriptions. We constructed a massive training corpus of 1.9 million photographs covering 695,091~unique animals to support this investigation. Through systematic ablation studies, we identified SigLIP2-Giant and E5-Small-v2 as the optimal vision and text backbones. We further evaluated fusion strategies ranging from simple concatenation to adaptive gating to determine the best method for integrating these modalities. Our proposed approach utilizes a gated fusion mechanism and achieved a Top-1 accuracy of 84.28% and an Equal Error Rate of 0.0422 on a comprehensive test protocol. These results represent an 11% improvement over leading unimodal baselines and demonstrate that integrating synthesized semantic descriptions significantly refines decision boundaries in large-scale pet re-identification.\\ Conference update. This work was accepted to the FGVC13 Workshop at CVPR 2026. We thank Vadim Ezhov and Nikolai Shcheglov for their help with the poster presentation and demonstration at the conference.

Figures & tables

Gap	Representative Works	How This Work Addresses It
No unified evaluation protocol	( Mougeot et al., 2019 ; Zhang et al., 2024 ; Kang and Oh, 2025 ; Vidal et al., 2021 )	We combine five large training datasets into a unified corpus and define a single standardized training and evaluation protocol used for all backbones and configurations.
Unimodality	( Mougeot et al., 2019 ; Cermak et al., 2025 ; Kang and Oh, 2025 ; Vidal et al., 2021 )	We introduce a multimodal pipeline that augments visual embeddings with identity-level textual descriptions.
No systematic ablation studies	( Mougeot et al., 2019 ; Zhang et al., 2024 ; Cermak et al., 2025 ; Kang and Oh, 2025 ; Vidal et al., 2021 )	We perform controlled ablations over multiple vision encoders, text encoders, and training configurations, isolating their contributions to verification performance on a common benchmark suite.
No controlled comparison of multimodal fusion strategies	( Mougeot et al., 2019 ; Zhang et al., 2024 ; Kang and Oh, 2025 ; Li et al., 2025 ; Vidal et al., 2021 )	We implement and compare several fusion mechanisms under identical settings, identifying effective designs for animal verification.

Table 1: Summary of key gaps in prior work and how the present study addresses them.

Dataset	Identities	Photos	Usage	Min	Max	Mean	Med	Std
Pet911 (ours)	22,050	65,961	Training	2	10	2.99	3	1.34
Telegram (ours)	73,101	131,698	Training	2	10	2.73	2	1.05
Dogs-World ( Toumbourou, 2022 )	200,458	301,342	Training	2	17	2.36	2	0.63
LCW ( DSEIDLI, 2023 )	140,732	381,267	Training	2	6	3.71	2	1.45
PetFace ( Shinoda and Shiohara, 2024 )	257,349	1,001,532	Training	3	11	4.58	4	1.29
Cat Individual Images ( Lin, 2019 )	518	13,542	Evaluation	6	144	25.58	20	10.06

Table 2: Dataset Composition and Statistics.

Configuration	Parameters	Mult-Adds	VRAM	Epoch Time	Throughput
CLIP-ViT-Base	151,277,313	201,094,656	375.24	7641	211
DINOv2-Small	22,056,576	79,195,392	201.66	2497	224
SigLIP-Base	203,155,970	205,686,528	500.96	4354	151
SigLIP2-Base	375,187,970	205,686,528	500.96	4271	160
Zer0int CLIP-L	427,616,513	457,503,744	1696.21	10,041	66
SigLIP2-Giant	1,871,885,426	1,833,395,712	7394.99	64,641	10

Table 3: Vision Encoders’ computational characteristics.

Configuration	Parameters	Mult-Adds	VRAM	Epoch Time	Throughput
BERT	109,482,240	109,482,240	519.65	669	990
E5-Base	109,482,240	109,482,240	519.65	688	985
E5-Base-v2	109,482,240	109,482,240	519.65	670	963
E5-Small	33,360,000	33,360,000	178.26	657	986
E5-Small-v2	33,360,000	33,360,000	178.26	661	997

Table 4: Text encoders’ computational characteristics.

Configuration	Parameters	Mult-Adds	VRAM	Epoch Time	Throughput
CLIP-ViT-Base + E5-Base-v2	261,415,937	311,233,280	897.41	2813	203
CLIP-ViT-Base + E5-Small-v2	185,097,089	234,914,432	555.27	2788	204
CLIP-ViT-Base + E5-Small-v2 + cross-attention	187,988,865	235,704,960	588.3	2952	195
CLIP-ViT-Base + E5-Small-v2 + weighted text	185,097,090	234,914,432	555.27	2905	194
SigLIP2-Giant + E5-Small-v2	1,907,803,890	1,869,314,176	7557.53	63,959	10
SigLIP2-Giant + E5-Small-v2 + cross-attention	1,907,805,426	1,868,265,088	7556.02	63,296	10

Table 5: Multimodal setups’ computational characteristics.

Configuration	Parameters	Mult-Adds	VRAM	Throughput
MiewID-msv3	51,109,277	24,307,036,098	891.97	159
MD-T-CNN-288	12,233,232	1,589,219,176	211.22	353
MD-CLIP-336	303,507,456	649,129,984	2360.4	36
MD-L-384	195,198,516	445,579,776	1960.17	46
BioCLIP	86,192,640	172,314,624	391.57	267

Table 6: Comparison methods’ computational characteristics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
Train data	0.9557	0.1042	0.6392	0.8088	0.8535
Train data + YOLO Preprocessing	0.9626	0.0978	0.6282	0.7826	0.8256
Train data + PetFace(train split)	0.9598	0.0982	0.6460	0.8093	0.8504
Train data + Telegram	0.9716	0.0818	0.6791	0.8329	0.8713
Train data + Pet911 + Telegram	0.9729	0.0772	0.6527	0.8139	0.8568
Train data + PetFace (full) + Pet911 + Telegram	0.9752	0.0729	0.6511	0.8122	0.8555

Table 7: Performance comparison for ablations across different data setups using the CLIP-ViT-Base model ( Radford et al., 2021 ) ; all metrics are computed on the entire test set.

Figure 1: Pet Identification Performance Across Training Data Configurations: ROC AUC, 1 - EER, and Top-k Metrics.

Configuration	Train Data	Train Data + YOLO Preprocessing	Train Data + PetFace (Train Split)	Train Data + Telegram	Train Data + Pet911 + Telegram
Train data + YOLO Preprocessing	0.0 (↑)	-	-	-	-
Train data + PetFace (train split)	0.000012 (↑)	0.593305 (↑)	-	-	-
Train data + Telegram	0.0 (↑)	0.0 (↑)	0.0 (↑)	-	-
Train data + Pet911 + Telegram	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.000152 (↑)	-
Train data + PetFace (full) + Pet911 + Telegram	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.001651 (↑)

Table 8: Pairwise McNemar test

p

-values for comparing training-data configurations. ↑ indicates fewer errors for the row encoder; ↓ indicates fewer errors for the column encoder.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
Train data	0.9753	0.0761	0.8082	0.9475	0.9655
Train data + YOLO Preprocessing	0.9783	0.0708	0.8411	0.9601	0.9721
Train data + PetFace(train split)	0.9772	0.0713	0.8491	0.9677	0.9790
Train data + Telegram	0.9857	0.0528	0.8575	0.9737	0.9832
Train data + Pet911 + Telegram	0.9822	0.0593	0.8296	0.9561	0.9712
Train data + PetFace (full) + Pet911 + Telegram	0.9821	0.0604	0.8359	0.9579	0.9711

Table 9: Performance comparison for ablations across different data setups using the CLIP-ViT-Base model ( Radford et al., 2021 ) ; all metrics are computed on the Cat Individual Images Lin (2019) .

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
Train data	0.9653	0.0920	0.4408	0.6477	0.7241
Train data + YOLO Preprocessing	0.9494	0.1207	0.3802	0.5772	0.6575
Train data + PetFace(train split)	0.9735	0.0774	0.4098	0.6264	0.7029
Train data + Telegram	0.9715	0.0866	0.4705	0.6688	0.7415
Train data + Pet911 + Telegram	0.9727	0.0778	0.4460	0.6479	0.7239
Train data + PetFace (full) + Pet911 + Telegram	0.9739	0.0772	0.4350	0.6417	0.7204

Table 10: Performance comparison for ablations across different data setups using the CLIP-ViT-Base model ( Radford et al., 2021 ) ; all metrics are computed on the DogFaceNet dataset Mougeot et al. (2019) .

Figure 2: t-SNE visualization of training data embeddings under different data configurations.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9752	0.0729	0.6511	0.8122	0.8555
DINOv2-Small	0.9848	0.0546	0.7180	0.8678	0.9009
SigLIP-Base	0.9811	0.0572	0.7359	0.8831	0.9140
SigLIP2-Base	0.9793	0.0631	0.7400	0.8889	0.9197
Zer0int CLIP-L	0.9842	0.0565	0.7626	0.8994	0.9267
SigLIP2-Giant	0.9912	0.0378	0.8243	0.9471	0.9641

Table 11: Vision Encoder Performance—Overall Metrics.

Figure 3: Pet Identification Performance Across Vision Encoder Configurations: ROC AUC, 1 - EER, and Top-k Metrics.

Configuration	CLIP-ViT-Base	DINOv2-Small	SigLIP-Base	SigLIP2-Base	Zer0int CLIP-L
DINOv2-Small	0.0 (↑)	-	-	-	-
SigLIP-Base	0.0 (↑)	0.005463 (↓)	-	-	-
SigLIP2-Base	0.0 (↑)	0.0 (↓)	0.0 (↓)	-	-
Zer0int CLIP-L	0.0 (↑)	0.134735 (↓)	0.30085 (↓)	0.0 (↓)	-
SigLIP2-Giant	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)

Table 12: Pairwise McNemar test

p

-values for comparing vision encoder configurations. ↑ indicates fewer errors for the row encoder; ↓ indicates fewer errors for the column encoder.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9821	0.0604	0.8359	0.9579	0.9711
DINOv2-Small	0.9904	0.0422	0.8547	0.9660	0.9764
SigLIP-Base	0.9899	0.0390	0.8649	0.9757	0.9842
SigLIP2-Base	0.9894	0.0388	0.8660	0.9772	0.9863
Zer0int CLIP-L	0.9881	0.0509	0.8768	0.9767	0.9845
SigLIP2-Giant	0.9940	0.0344	0.8899	0.9868	0.9921

Table 13: Vision Encoder Performance—Cat Individual Images Lin (2019) Metrics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9739	0.0772	0.4350	0.6417	0.7204
DINOv2-Small	0.9829	0.0571	0.5581	0.7540	0.8139
SigLIP-Base	0.9792	0.0606	0.5848	0.7746	0.8319
SigLIP2-Base	0.9776	0.0672	0.5925	0.7856	0.8422
Zer0int CLIP-L	0.9814	0.0625	0.6289	0.8092	0.8597
SigLIP2-Giant	0.9926	0.0326	0.7475	0.9009	0.9316

Table 14: Vision Encoder Performance—DogFaceNet Mougeot et al. (2019) Metrics.

Figure 4: t-SNE visualization of embeddings under different vision encoder configurations.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
BERT	0.9698	0.0736	0.3675	0.5717	0.6554
E5-Base	0.9703	0.0740	0.3733	0.5748	0.6595
E5-Base-v2	0.9697	0.0755	0.3699	0.5743	0.6575
E5-Small	0.9693	0.0764	0.3795	0.5783	0.6615
E5-Small-v2	0.9700	0.0757	0.3736	0.5731	0.6597

Table 15: Text Encoder Performance—Overall Metrics.

Configuration	BERT	E5-Base	E5-Base-v2	E5-Small
E5-Base	0.696150 (↓)	-	-	-
E5-Base-v2	0.013254 (↓)	0.029675 (↓)	-	-
E5-Small	0.044146 (↑)	0.093004 (↓)	0.771638 (↑)	-
E5-Small-v2	0.016518 (↑)	0.041364 (↓)	1.000000 (↓)	0.731556 (↓)

Table 16: Pairwise McNemar test

p

-values for comparing text encoder configurations. ↑ indicates fewer errors for the row encoder; ↓ indicates fewer errors for the column encoder.

Figure 5: Pet Identification Performance Across Text Encoder Configurations: ROC AUC, 1 - EER, and Top-k Metrics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
BERT	0.9707	0.0726	0.5301	0.7397	0.8060
E5-Base	0.9706	0.0739	0.5396	0.7464	0.8157
E5-Base-v2	0.9708	0.0751	0.5386	0.7484	0.8131
E5-Small	0.9693	0.0764	0.5510	0.7504	0.8188
E5-Small-v2	0.9700	0.0754	0.5373	0.7439	0.8146

Table 17: Text Encoder Performance—Cat Individual Images Lin (2019) Metrics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
BERT	0.9756	0.0674	0.1776	0.3745	0.4791
E5-Base	0.9758	0.0674	0.1789	0.3731	0.4765
E5-Base-v2	0.9759	0.0676	0.1721	0.3713	0.4751
E5-Small	0.9745	0.0685	0.1804	0.3761	0.4772
E5-Small-v2	0.9753	0.0692	0.1821	0.3743	0.4788

Table 18: Text Encoder Performance—DogFaceNet Mougeot et al. (2019) Metrics.

Figure 6: t-SNE visualization of embeddings under different text encoder configurations.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9752	0.0729	0.6511	0.8122	0.8555
CLIP-ViT-Base + E5-Base-v2	0.9711	0.0768	0.6331	0.8130	0.8561
CLIP-ViT-Base + E5-Small-v2	0.9707	0.0800	0.6454	0.8193	0.8617
CLIP-ViT-Base + E5-Small-v2 + cross-attention	0.9807	0.0627	0.6409	0.8149	0.8589
CLIP-ViT-Base + E5-Small-v2 + weighted text	0.9703	0.0793	0.6696	0.8371	0.8806
SigLIP2-Giant	0.9912	0.0378	0.8243	0.9471	0.9641

Table 19: Multimodal setups Performance—Overall Metrics.

Figure 7: Pet Identification Performance Across Multimodal Configurations: ROC AUC, 1 - EER, and Top-k Metrics.

Configuration	CLIP-ViT-Base + E5-Base-v2	CLIP-ViT-Base + E5-Small-v2	CLIP-ViT-Base + E5-Small-v2 + Cross-Attention	CLIP-ViT-Base + E5-Small-v2 + Weighted Text	SigLIP2-Giant + E5-Small-v2	SigLIP2-Giant + E5-Small-v2 + Cross-Attention	SigLIP2-Giant + E5-Small-v2 + Weighted Text
CLIP-ViT-Base + E5-Small-v2	0.004174 (↓)	-	-	-	-	-	-
CLIP-ViT-Base + E5-Small-v2 + cross-attention	0.0 (↓)	0.0 (↓)	-	-	-	-	-
CLIP-ViT-Base + E5-Small-v2 + weighted text	0.044237 (↓)	0.524457 (↑)	0.0 (↑)	-	-	-	-
SigLIP2-Giant + E5-Small-v2	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	-	-	-
SigLIP2-Giant + E5-Small-v2 + cross-attention	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.002496 (↑)	-	-
SigLIP2-Giant + E5-Small-v2 + weighted text	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	-

Table 20: Pairwise McNemar test

p

-values for comparing mulimodal models configurations. ↑ indicates fewer errors for the row encoder; ↓ for the column encoder.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9821	0.0604	0.8359	0.9579	0.9711
CLIP-ViT-Base + E5-Base-v2	0.9824	0.0617	0.8004	0.9445	0.9634
CLIP-ViT-Base + E5-Small-v2	0.9839	0.0596	0.8141	0.9532	0.9694
CLIP-ViT-Base + E5-Small-v2 + cross-attention	0.9807	0.0627	0.8062	0.9477	0.9669
CLIP-ViT-Base + E5-Small-v2 + weighted text	0.9843	0.0587	0.8176	0.9497	0.9693
SigLIP2-Giant	0.9940	0.0344	0.8899	0.9868	0.9921

Table 21: Multimodal setups Performance—Cat Individual Images Lin (2019) Metrics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
CLIP-ViT-Base	0.9739	0.0772	0.4350	0.6417	0.7204
CLIP-ViT-Base + E5-Base-v2	0.9739	0.0744	0.4371	0.6589	0.7309
CLIP-ViT-Base + E5-Small-v2	0.9717	0.0817	0.4479	0.6628	0.7358
CLIP-ViT-Base + E5-Small-v2 + cross-attention	0.9715	0.0836	0.4469	0.6593	0.7326
CLIP-ViT-Base + E5-Small-v2 + weighted text	0.9786	0.0676	0.4960	0.7054	0.7768
SigLIP2-Giant	0.9926	0.0326	0.7475	0.9009	0.9316

Table 22: Multimodal setups Performance—DogFaceNet Mougeot et al. (2019) Metrics.

Figure 8: t-SNE visualization of embeddings under different multimodal setup configurations.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
MiewID-msv3	0.8484	0.2442	0.7270	0.8468	0.8685
MiewID-msv3	0.9227	0.1508	0.5569	0.7057	0.7447
MD-T-CNN-288	0.7592	0.3082	0.1599	0.2515	0.2955
MD-CLIP-336	0.7767	0.2998	0.3410	0.4648	0.5138
MD-L-384	0.9070	0.1673	0.4607	0.6170	0.6695
BioCLIP	0.8383	0.2465	0.3720	0.5047	0.5539

Table 23: Comparison of the proposed methods with existing approaches—Overall Metrics.

Figure 9: Pet Identification Performance Across Compared Approaches: ROC AUC, 1 - EER, and Top-k Metrics.

Configuration	MiewID-msv3	MD-T-CNN-288	MD-CLIP-336	MD-L-384	BioCLIP
MD-T-CNN-288	0.0 (↓)	-	-	-	-
MD-CLIP-336	0.0 (↓)	0.0 (↑)	-	-	-
MD-L-384	0.0 (↓)	0.0 (↓)	0.0 (↑)	-	-
BioCLIP	0.0 (↓)	0.0 (↓)	0.0 (↓)	0.0 (↓)	-
SigLIP2-Giant + E5-Small-v2 + gating	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)	0.0 (↑)

Table 24: Pairwise McNemar test

p

-values for comparing existing approaches. ↑ indicates fewer errors for the row encoder; ↓ indicates fewer errors for the column encoder.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
MiewID-msv3	0.9383	0.1392	0.8723	0.9695	0.9760
MD-T-CNN-288	0.8695	0.2111	0.7347	0.8546	0.8743
MD-CLIP-336	0.8953	0.1848	0.8182	0.9158	0.9277
MD-L-384	0.9364	0.1401	0.8421	0.9464	0.9567
BioCLIP	0.8754	0.2141	0.8473	0.9478	0.9585
SigLIP2-Giant + E5-Small-v2 + gating	0.9929	0.0344	0.8952	0.9872	0.9932

Table 25: Comparison of the proposed methods with existing approaches—Cat Individual Images Lin (2019) Metrics.

Configuration	ROC AUC	EER	Top-1	Top-5	Top-10
MiewID-msv3	0.9227	0.1508	0.5569	0.7057	0.7447
MD-T-CNN-288	0.7592	0.3082	0.1599	0.2515	0.2955
MD-CLIP-336	0.7767	0.2998	0.3410	0.4648	0.5138
MD-L-384	0.9070	0.1673	0.4607	0.6170	0.6695
BioCLIP	0.8383	0.2465	0.3720	0.5047	0.5539
SigLIP2-Giant + E5-Small-v2 + gating	0.9920	0.0314	0.7818	0.9233	0.9482

Table 26: Comparison of the proposed methods with existing approaches—DogFaceNet Mougeot et al. (2019) Metrics.

Figure 10: t-SNE visualization of embeddings under different feature extraction models, comparing the proposed method against existing approaches.

Animal Re-Identification

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

See all 33 papers

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
Advancing Wildlife Conservation through Multimodal Animal Re-Identification with Environmental Metadata
Sep 28, 2026Yuzhuo Li, Di Zhao, Tingrui Qiao +3Animal Re-Identification Wildlife Monitoring
School of Computer Science, University of Auckland Auckland, New Zealand
Visual-Prompt Guided Wildlife Instance-Level Recognition
Aug 18, 2026Mufhumudzi Muthivhi, Jiahao Huo, Terence van Zyl +1Animal Re-Identification Object Detection
University of Johannesburg, South Africa · Linköping University, Sweden

Jul 17, 2026J/K move · Enter open · S save

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

Evan Sinclair Smith, Anthony Miyaguchi, Snigdha Palamari, Danté Evangelista

1Georgia Institute of Technology, North Ave NW, Atlanta, GA 30332

Animal Re-Identification Image Matching Graph Clustering Wildlife Monitoring

Open paper AskPDF arXiv

Automated individual animal re-identification is essential for large-scale biodiversity monitoring; however, field imagery complicates separating identity cues from nuisance variation in pose, illumination, background, resolution, and species-specific morphology. The DS@GT ARC submission to AnimalCLEF 2026 introduces a multi-species image-clustering system for re-identifying Eurasian lynx, fire salamanders, loggerhead sea turtles, and Texas horned lizards. Instead of relying on a single descriptor or nearest-neighbor retrieval, this approach formulates re-identification as species-aware graph construction over candidate image pairs. The pipeline integrates tailored preprocessing, global candidate retrieval, LightGlue-based local verification with multiple keypoint families, LightGBM pair scoring, conservative edge admission, and Leiden community detection. This design directly addresses a primary failure mode of clustering-based re-identification: high-scoring false pairs that act as bridge edges and merge distinct individuals through transitive closure. Across species, ablation studies demonstrate that local feature support, foreground-aware preprocessing, and species-specific backbone selection enhance pair evidence, while graph operating points determine the trade-off between fragmentation and over-merging. The selected submission achieved a public ARI of 0.733 and a private ARI of 0.674, ranking fifth among 230 teams. These results indicate that robust wildlife re-identification requires not only strong visual representations but also calibrated integration of global similarity, local identity markings, neighborhood context, and graph-level constraints. The code can be found at https://github.com/dsgt-arc/animalclef-2026.

Abstract

Animal Re-Identification

Explore similar work

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

Advancing Wildlife Conservation through Multimodal Animal Re-Identification with Environmental Metadata

Visual-Prompt Guided Wildlife Instance-Level Recognition