Are We Recognizing the Jaguar or Its Background? A Diagnostic Framework for Jaguar Re-Identification

Antonio Rueda-Toicen; Abigail Allen Martin; Daniil Morozov; Matin Mahmood; Alexandra Schild; Shahabeddin Dayani; Davide Panza; Gerard de Melo

cs.CVApr 6, 2026

Are We Recognizing the Jaguar or Its Background? A Diagnostic Framework for Jaguar Re-Identification

Authors: Antonio Rueda-Toicen, Abigail Allen Martin, Daniil Morozov, Matin Mahmood, Alexandra Schild, Shahabeddin Dayani, Davide Panza, Gerard de Melo

Organizations: Hasso Plattner Institute · NVIDIA · Jaguar ID Project · Universidade Federal de Mato Grosso ·

Abstract

Jaguar re-identification (re-ID) from citizen-science imagery can look strong on standard retrieval metrics while still relying on the wrong evidence, such as background context or silhouette shape, instead of the coat pattern that defines identity. We introduce a diagnostic framework for wildlife re-ID with two axes: a leakage-controlled context ratio, background/foreground, computed from inpainted background-only versus foreground-only images, and a laterality diagnostic based on cross-flank retrieval and mirror self-similarity. To make these diagnostics measurable, we curate a Pantanal jaguar benchmark with per-pixel segmentation masks and an identity-balanced evaluation protocol. We then use representative mitigation families, ArcFace fine-tuning, anti-symmetry regularization, and Lorentz hyperbolic embeddings, as case studies under the same evaluation lens. The goal is not only to ask which model ranks best, but also what visual evidence it uses to do so.

Figures & tables

Figure 1: Six diagnostic image variants derived from a single citizen-science photograph using its SAM-3 alpha mask. (A1) Original full_rgb input. (A2) Binary silhouette (alpha channel). (B1) bg_silhouette : foreground replaced by a black silhouette, retaining background context and shape cues. (B2) inpainted : foreground removed by FLUX.1-Fill generative inpainting, eliminating the silhouette-shaped hole used for leakage-controlled BG/FG evaluation ( Section 3.2 ). (C1) foreground cutout (background removed), isolating coat pattern. (C2) Horizontally flipped mirror cutout, used for laterality analysis ( Section 3.3 ). Jaguar rosette patterns are individually unique [ 1 ] ; the left flank differs from the right, so a mirrored image depicts an individual that does not exist.

Rank	Model	mAP (macro)	BG/FG	Mirror Sim	Mirror Tier	Architecture
1	MiewID-MSv3	0.303	0.59	0.746	T1: Aware	EffNetV2-M
2	DINOv3-ViT CLS	0.298	0.77	0.963	T2: Moderate	ViT-S/16
3	MiewID-MSv2	0.297	0.52	0.752	T1: Aware	EffNetV2-M
4	HotSpotter+VLAD	0.288	0.54	1.000	T4: Symmetric	SIFT+VLAD
5	EVA-02	0.281	0.66	0.997	T4: Symmetric	ViT-L/14
6	DINOv3-ViT Pooled	0.277	0.82	0.979	T3: Strong	ViT-S/16

Table 1: Frozen baseline benchmark. 16 models evaluated on three variants ( full_rgb , foreground_only , background_inpainted ), yielding 48 test-set evaluations. The mAP column is identity-balanced (macro) test mAP on full_rgb . BG/FG uses background_inpainted / foreground_only (Eq. 3 ); mirror similarity is mean mirror_sim on foreground_only (Eq. 4 ). Models are ranked by mAP.

Figure 2: Foreground vs. inpainted background mAP across frozen models. Green bars show foreground-only mAP; red bars show inpainted background-only mAP. Models are sorted by BG/FG , shown in bold next to each bar pair. MiewID-MSv2 achieves the lowest frozen BG/FG (0.52), consistent with wildlife-specific pre-training reducing non-coat context reliance.

Figure 3: Relationship between shortcut axes. Each point is a frozen model plotted by inpainted BG/FG (

x

) and mean mirror similarity (

y

), both computed on foreground-only cutouts (lower mean mirror similarity indicates greater laterality awareness). Spearman

\rho=0.307

(

p=0.265

;

N{=}15

; 95% bootstrap CI

[-0.360,\,0.771]

B{=}20{,}000

, seed 0): no clear monotonic association. Counter-examples include EVA-02 ( BG/FG

0.661

, mean mirror similarity

0.997

) and I-JEPA ( BG/FG

1.166

, mean mirror similarity

0.968

	Within-flank		Cross-flank		Cross/Within	Discrim-
Model	mAP	CMC@1	mAP	CMC@1	ratio	inability
ArcFace+MiewID (no flip)	0.716	77.5%	0.477	53.9%	0.666	0.417
ArcFace+MiewID (flip aug)	0.632	70.7%	0.490	55.3%	0.774	0.427
$\Delta$ (flip $-$ no flip)	$-$ 0.084	$-$ 6.8pp	+0.013	+1.4pp	+0.108	+0.010

Table 2: Cross-flank retrieval ablation (foreground-only, macro metrics). Flip augmentation does not improve cross-flank mAP and reduces within-flank accuracy by 12%.

	Full RGB				Cross-Flank
Model	mAP	CMC@1	CMC@5	CMC@10	mAP	CMC@1	CMC@5	CMC@10
ArcFace O0 (MSv3)	0.483 $\pm$ 0.011	74.3% $\pm$ 0.9%	80.7% $\pm$ 1.0%	83.8% $\pm$ 0.4%	0.576 $\pm$ 0.049	60.7% $\pm$ 4.9%	69.0% $\pm$ 4.3%	75.1% $\pm$ 2.6%
ArcFace O1 (MSv3)	0.469 $\pm$ 0.010	74.3% $\pm$ 0.9%	80.4% $\pm$ 1.5%	82.4% $\pm$ 1.0%	0.566 $\pm$ 0.026	62.4% $\pm$ 1.8%	69.9% $\pm$ 4.7%	76.2% $\pm$ 5.6%
Lorentz O0 (MSv3)	0.548 $\pm$ 0.012	76.3% $\pm$ 1.5%	83.1% $\pm$ 1.0%	86.3% $\pm$ 0.7%	0.689 $\pm$ 0.032	72.0% $\pm$ 4.7%	81.5% $\pm$ 0.7%	84.8% $\pm$ 2.5%
Lorentz O1 (MSv2)	0.515 $\pm$ 0.014	75.0% $\pm$ 1.7%	83.5% $\pm$ 0.3%	87.6% $\pm$ 0.9%	0.653 $\pm$ 0.011	71.2% $\pm$ 4.9%	84.4% $\pm$ 1.2%	88.8% $\pm$ 1.0%
Lorentz O1 (MSv3)	0.557 $\pm$ 0.009	77.3% $\pm$ 1.7%	84.3% $\pm$ 1.4%	87.1% $\pm$ 1.4%	0.692 $\pm$ 0.010	72.5% $\pm$ 2.2%	81.9% $\pm$ 3.4%	87.0% $\pm$ 2.9%

Table 3: TEST-only retrieval summary for the matched objective-pair benchmark. Values are family-level mean

\pm

std over seeds 42/43/44. Full-RGB and cross-flank endpoints both report macro mAP and macro CMC@1/5/10. These rows describe benchmark ordering; pairwise analyses appear on the Appendix’s Wilcoxon tables.

Panel of Figure 4 — Figure 4: Qualitative embedding inspection with UMAP [ 28 ] in HyperView. The interface shows image thumbnails next to the latent arrangement and includes a lasso tool for selecting regions for qualitative inspection, near-duplicate discovery, and difficult re-identification cases. Hard cases produce overlapping embedding representations. These panels are qualitative explorations and part of the iterative model-guided dataset curation process.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Backbone	Architecture	Pre-training	Dim	Frozen	Fine-tuned
Wildlife-specific pre-training
MiewID-MSv2 [ 6 ]	EfficientNetV2-M	Wildlife (49 spp.)	2152	$\checkmark$	A, T, L, AS
MiewID-MSv3 [ 6 ]	SwinV2	Wildlife	2152	$\checkmark$	A, T, L
MegaDescriptor-L [ 4 ]	Swin-L	Wildlife	1536	$\checkmark$	A, T
Self-supervised / multi-task
DINOv3-ViT [ 10 ]	ViT-S/16	Self-sup. (LVD-142M)	384	$\checkmark^{\dagger}$	A

Appendix

Table 4: All 16 backbone architectures evaluated. Dim = frozen embedding dimension before any projection head. Frozen = included in the frozen baseline evaluation (Table 1 in the main paper). Fine-tuning abbreviations: A = ArcFace, T = Triplet, L = Lorentz, AS = Anti-symmetry.

Family / model	Training input	Size	Checkpoint selection	Seed handling	Final statistic	Representative checkpoint rule
ArcFace+MiewID	FG cutouts	440	Lowest val loss on coreset	Seeds 42/43/44	Test mean $\pm$ std	Lowest val-loss seed
ArcFace+MiewID+AntiSym ( $\tau{=}0.3,\lambda{=}0.5$ )	FG cutouts	440	Lowest val loss on coreset	Seeds 42/43/44; hyperparams fixed before reruns	Test mean $\pm$ std	Lowest val-loss seed
Lorentz+MSv3+O0	FG cutouts + solidity	320	Family: mean val ib-mAP; seed ckpt: val loss	Seeds 42/43/44	Test mean $\pm$ std	Highest val ib-mAP, lower val loss tiebreak
Lorentz+MSv2+O1	FG cutouts + solidity	320	Family: mean val ib-mAP; seed ckpt: val loss	Seeds 42/43/44	Test mean $\pm$ std	Highest val ib-mAP, lower val loss tiebreak
Lorentz+MSv3+O1	FG cutouts + solidity	320	Family: mean val ib-mAP; seed ckpt: val loss	Seeds 42/43/44	Test mean $\pm$ std	Highest val ib-mAP, lower val loss tiebreak
ArcFace+DINOv3	FG cutouts	224	Lowest val loss on coreset	Seeds 42/43/44	Test mean $\pm$ std	Lowest val-loss seed

Appendix

Table 5: Explicit comparison protocol for the trainable families referenced in the main paper and mitigation supplement. “FG cutouts” means the alpha-mask foreground cutout is applied at load time. Lorentz family selection is first done at the family level by mean validation ib-mAP across seeds; the representative checkpoint is then chosen on validation statistics only.

Family	Seed	Checkpoint	Full-RGB mAP	Full-RGB CMC@1	Cross-Flank mAP	Cross-Flank CMC@1	Raw mirror
ArcFace O0 (MSv3)	42	arcface_miewid_msv3_O0_seed42_best.pth	0.475	73.2%	0.582	65.1%	0.858
ArcFace O0 (MSv3)	43	arcface_miewid_msv3_O0_seed43_best.pth	0.476	75.5%	0.513	53.8%	0.850
ArcFace O0 (MSv3)	44	arcface_miewid_msv3_O0_seed44_best.pth	0.499	74.2%	0.633	63.1%	0.857
ArcFace O1 (MSv3)	42	arcface_miewid_msv3_O1_seed42_best.pth	0.479	75.4%	0.576	64.8%	0.849
ArcFace O1 (MSv3)	43	arcface_miewid_msv3_O1_seed43_best.pth	0.473	74.4%	0.530	60.4%	0.844
ArcFace O1 (MSv3)	44	arcface_miewid_msv3_O1_seed44_best.pth	0.456	73.1%	0.591	61.9%	0.848

Appendix

Table 6: Per-seed TEST metrics for the six matched objective-pair families in the main benchmark table (seeds 42/43/44). We report full-RGB macro mAP, full-RGB macro CMC@1, cross-flank macro mAP, cross-flank macro CMC@1, and raw Axis 2 mirror similarity, together with the checkpoint retained for traceability.

Split	$n$	Mean	Median	Min	Max	Frac( $s<0.5$ )
Train	1,895	0.6937	0.7094	0.2277	0.9678	0.0770
Test	371	0.6883	0.6967	0.2388	0.9561	0.0701

Appendix

Table 7: Mask solidity statistics computed from the SAM 3 alpha masks used in our experiments (Eq. 8 ). “Frac(

s<0.5

)” is the fraction of images with solidity below 0.5.

Figure 5: Mask solidity distribution for the train ( $n{=}1{,}895$ ) and test ( $n{=}371$ ) splits, computed from SAM 3 alpha masks (Eq. 8 ).

Figure 6: Lowest-solidity masks. Examples from the bottom of the solidity distribution. The most common artifact we see is partial occlusion (often vegetation), which removes parts of the jaguar or introduces small holes.

Frozen baselines
Model	Full mAP	FG mAP	Inpainted mAP	BG/FG	BG+Sil/FG	Risk
MiewID-MSv2	0.297	0.292	0.151	0.516	0.848	LOW
HotSpotter+VLAD	0.288	0.271	0.145	0.535	0.846	LOW
MiewID-MSv3	0.303	0.304	0.178	0.586	0.869	LOW
EVA-02	0.281	0.303	0.200	0.661	0.940	LOW
DINOv3-ViT CLS	0.298	0.296	0.228	0.768	1.020	LOW

Appendix

Table 8: Leakage-controlled inpainted BG/FG results across frozen baselines and mitigation finalists (test split). Full mAP, FG mAP, and Inpainted mAP are identity-balanced (macro), computed on full_rgb , foreground_only , and background_inpainted , respectively. Models are grouped by type and sorted by BG/FG .

Metric	Mean	Median	Min	Max	# ${\geq}1$	# ${>}1.1$
Legacy BG+Sil/FG	0.501	0.440	0.172	1.355	12	7
Inpainted BG/FG	0.396	0.324	0.120	1.166	3	2
Silhouette/FG	0.366	0.348	0.125	0.860	0	0

Appendix

Table 9: Summary statistics across all 102 evaluated models. Legacy BG+Sil/FG uses the naïve background_only construction; inpainted BG/FG uses the leakage-controlled background_inpainted variant.

Tier	Negative Type	Defeats	Origin
1 (Hard)	Mirror (LR flip)	Laterality shortcut	Axis 2
1 (Hard)	Background cutout	Non-coat context memorisation	Axis 1
2 (Semi-hard)	Closest different individual	Confusable pairs	Offline ranking

Appendix

Table 10: Three-tier negative mining strategy linking diagnostics to training.

Backbone	O0 (standard)	O1 (mirror neg)	$\Delta$
MiewID-MSv3	$0.537\pm 0.034$	$\mathbf{0.608\pm 0.007}$	$+$ 13.3%
MiewID-MSv2	$0.532\pm 0.002$	$\mathbf{0.565\pm 0.024}$	$+$ 6.0%
ConvNeXt-tiny	$0.326\pm 0.003$	$\mathbf{0.328\pm 0.008}$	$+$ 0.6%
Swin-tiny	$\mathbf{0.312\pm 0.006}$	$0.310\pm 0.005$	$-$ 0.7%
EfficientNetV2	$\mathbf{0.305\pm 0.005}$	$0.301\pm 0.006$	$-$ 1.2%
MegaDescriptor-L	$\mathbf{0.243\pm 0.008}$	$0.239\pm 0.005$	$-$ 1.6%

Appendix

Table 11: Lorentz hyperbolic benchmark (validation mAP, mean

\pm

std across 3 seeds). O1 adds mirror-negative loss. Bold: best per backbone.

Backbone	Frozen mAP	O0 (ArcFace)	O1 (ArcFace+AntiSym)	BG+Sil/FG (O0)
Swin-T	0.312	$0.469\pm 0.008$	$0.475\pm 0.008$	$0.455\pm 0.008$
ConvNeXt-T	0.326	$0.452\pm 0.003$	$0.450\pm 0.016$	$0.443\pm 0.014$
EffNetV2-M	0.193	$0.452\pm 0.014$	$0.450\pm 0.015$	$0.341\pm 0.005$
MegaDesc-L	0.245	$0.378\pm 0.009$	$0.363\pm 0.016$	$0.433\pm 0.015$

Appendix

Table 12: ArcFace backbone benchmark (test macro mAP, mean

\pm

std across 3 seeds). O0: ArcFace only; O1: ArcFace + anti-symmetry (

\tau{=}0.5

\lambda{=}0.1

). All four backbones improve over their frozen baselines (gains range from

+

38.7% to

+

134.2%).

$\tau$	$\lambda$	mAP (macro)	CMC@1 (macro)	BG/FG
baseline		0.458	72.9%	0.314
0.3	0.1	0.497	75.1%	0.255
0.3	0.3	0.480	74.0%	0.244
0.3	0.5	0.498	74.9%	0.249
0.5	0.1	0.462	70.5%	0.249
0.5	0.3	0.480	72.2%	0.244

Appendix

Table 13: Anti-symmetry ablation (

3{\times}3

grid). Baseline is ArcFace+MiewID without

\mathcal{L}_{\text{anti}}

. Hyperparameters are selected on the validation coreset; test-set metrics are reported for transparency.

	T0 (Triplet only)			T1 (Triplet + AntiSym)
Backbone	Full mAP	FG mAP	BG/FG	Full mAP	FG mAP	BG/FG
MiewID-MSv3	$0.134\pm 0.014$	$0.549\pm 0.006$	$0.148\pm 0.012$	$\mathbf{0.510\pm 0.003}$	$0.574\pm 0.005$	$0.312\pm 0.009$
MiewID-MSv2	$0.088\pm 0.009$	$0.531\pm 0.007$	$0.136\pm 0.019$	$0.458\pm 0.008$	$0.554\pm 0.012$	$0.280\pm 0.005$
EffNetV2-M	$0.407\pm 0.003$	$0.495\pm 0.004$	$0.456\pm 0.008$	$0.348\pm 0.005$	$0.436\pm 0.001$	$0.511\pm 0.009$
Swin-T	$0.379\pm 0.008$	$0.493\pm 0.005$	$0.472\pm 0.004$	$0.338\pm 0.012$	$0.503\pm 0.006$	$0.485\pm 0.007$
ConvNeXt-T	$0.337\pm 0.007$	$0.502\pm 0.003$	$0.372\pm 0.003$	$0.322\pm 0.006$	$0.524\pm 0.007$	$0.370\pm 0.002$

Appendix

Table 14: Triplet loss benchmark (test macro mAP, mean

\pm

std across 3 seeds). T0: triplet only; T1: triplet + anti-symmetry (

\tau{=}0.5

\lambda{=}0.1

). FG mAP is foreground-only; BG/FG is the inpainted leakage-controlled ratio.

Model	X-mAP	X-CMC@1	X-CMC@5	X-CMC@10	X/W
ArcFace O0 (MSv3)	0.576 $\pm$ 0.049	60.7% $\pm$ 4.9%	69.0% $\pm$ 4.3%	75.1% $\pm$ 2.6%	0.777 $\pm$ 0.064
ArcFace O1 (MSv3)	0.566 $\pm$ 0.026	62.4% $\pm$ 1.8%	69.9% $\pm$ 4.7%	76.2% $\pm$ 5.6%	0.780 $\pm$ 0.044
Lorentz O0 (MSv3)	0.689 $\pm$ 0.032	72.0% $\pm$ 4.7%	81.5% $\pm$ 0.7%	84.8% $\pm$ 2.5%	0.899 $\pm$ 0.026
Lorentz O1 (MSv2)	0.653 $\pm$ 0.011	71.2% $\pm$ 4.9%	84.4% $\pm$ 1.2%	88.8% $\pm$ 1.0%	0.881 $\pm$ 0.028
Lorentz O1 (MSv3)	0.692 $\pm$ 0.010	72.5% $\pm$ 2.2%	81.9% $\pm$ 3.4%	87.0% $\pm$ 2.9%	0.876 $\pm$ 0.021
Triplet T0 (MSv3)	0.681 $\pm$ 0.013	73.0% $\pm$ 2.3%	81.3% $\pm$ 2.4%	84.1% $\pm$ 1.2%	0.923 $\pm$ 0.012

Appendix

Table 15: Cross-flank retrieval summary for the active v5 family-level bundle (foreground-only, macro metrics). Queries without a positive in the relevant gallery are excluded from that split. We report full cross-flank mAP and CMC@1/5/10 together with the cross/within ratio.

Comparison	Supported endpoints	Remaining tested endpoints
ArcFace O0 vs O1 (MSv3)	axis2 danger margin	All Axis-1 AP endpoints, raw mirror similarity, and cross-flank AP are not supported after Fisher combination and Holm correction.
Lorentz O0 vs O1 (MSv3)	none	Full-RGB AP, foreground AP, silhouette AP, inpainted AP, raw mirror similarity, danger margin, and cross-flank AP are all not supported.
Triplet T0 vs T1 (MSv3)	inpainted vegetation 1 AP; inpainted vegetation 2 AP; axis2 danger margin; axis2 mirror similarity	Full-RGB AP, foreground AP, silhouette AP, silhouette+bg AP, vegetation 3 AP, and cross-flank AP are not supported.

Appendix

Table 16: Paired AP and Axis 2 summary from the active v5 bundle. Rows use the matched-objective comparison policy: same family, same variant, different objective, complete seeds 42/43/44, two-sided Wilcoxon per seed, Fisher combination, and Holm correction.

Comparison	Supported endpoints	Note
ArcFace O0 vs O1 (MSv3)	none	All full-RGB, foreground, inpainted, silhouette, silhouette+bg, and cross-flank CMC@1/5/10 endpoints are not supported.
Lorentz O0 vs O1 (MSv3)	none	All full-RGB, foreground, inpainted, silhouette, silhouette+bg, and cross-flank CMC@1/5/10 endpoints are not supported.
Triplet T0 vs T1 (MSv3)	none	All full-RGB, foreground, inpainted, silhouette, silhouette+bg, and cross-flank CMC@1/5/10 endpoints are not supported.

Appendix

Table 17: Paired CMC summary from the same active v5 bundle. No matched objective pair has a supported CMC@1/5/10 difference after the full Wilcoxon-Fisher-Holm pipeline.

Method	mAP	BG/FG	BG+Sil/FG	Mirror Sim	Emb. Dim
HotSpotter+VLAD	0.288	0.535	0.845	1.000	8,192
HotSpotter+Mean	0.191	0.806	1.047	1.000	128

Appendix

Table 18: Classical CV baseline. HotSpotter+VLAD shows low shortcut reliance under both metrics, with zero learned parameters but laterality-blind.

Metric	Frozen DINOv3	ArcFace+DINOv3	Change
Test mAP (macro)	0.298	0.458	$+$ 53.7%
Test mAP (micro)	0.270	0.479	$+$ 77.4%
CMC@1 (macro)	70.0%	78.5%	$+$ 12.1%
CMC@1 (micro)	72.8%	82.2%	$+$ 12.9%
CMC@5 (macro)	77.6%	87.6%	$+$ 12.9%
CMC@5 (micro)	82.2%	89.8%	$+$ 9.2%

Appendix

Table 19: ArcFace+DINOv3-ViT-S/16 results. Self-supervised features match wildlife-specific features on mAP with superior CMC@1 and parameter efficiency.

Rank	Model	Mirror Sim	Std	BG+Sil/FG	Tier
1	MiewID-MSv3 (frozen)	0.746	0.067	NA	T1: Laterality-Aware
2	MiewID-MSv2 (frozen)	0.752	0.064	0.850	T1: Laterality-Aware
3	ArcFace+MiewID-MSv2	0.833	0.083	0.348	T1: Laterality-Aware
4	MegaDescriptor	0.914	0.035	1.020	T2: Moderate
5	ResNet50	0.958	0.021	1.125	T2: Moderate
6	DINOv3-ViT CLS	0.963	0.012	1.020	T2: Moderate

Appendix

Table 20: Mirror similarity ranking (15 models). Wildlife pre-training (MiewID) is the dominant factor; self-supervised models exceed 0.96. Standard deviations computed across 371 test images; T1 models show the highest per-image variance (std

>0.06

), indicating identity-dependent laterality sensitivity.

Correlation Metric	Value	$p$ -value	95% CI
Spearman $\rho$	$-0.200$	$0.493$	$[-0.757,\,0.404]$
Kendall $\tau$	$-0.187$	$0.388$	$[-0.643,\,0.284]$

Appendix

Table 21: Rank correlation between inpainted BG/FG and the mirrored-query retrieval ratio

r=\text{mAP}_{\text{mirror}}/\text{mAP}_{\text{regular}}

(foreground-only cutouts) across the

N{=}15

frozen models used in the main paper. This is an auxiliary retrieval stress test, not the canonical definition of Axis 2. Bootstrap 95% CIs are percentile intervals with

B{=}20{,}000

resamples (seed 0).

Tier	# IDs	% Data	CMC@1	CMC@5	CMC@10
HEAD (top 33%)	10	65%	86.2%	93.1%	95.4%
MEDIUM (middle 33%)	11	26%	80.0%	87.0%	89.0%
TAIL (bottom 33%)	10	9%	70.4%	81.5%	85.2%

Appendix

Table 22: Long-tail performance breakdown for ArcFace+DINOv3-ViT-S/16. Tail identities (9% of data) still achieve 70.4% CMC@1.

Figure 9: Long-tail performance breakdown for ArcFace+DINOv3. CMC@1/5/10 by identity frequency tier. Tail identities (10 IDs, 9% of data) still achieve 70.4% CMC@1, with a HEAD-to-TAIL gap of 15.8pp.

Figure 10: Auxiliary mirrored-query retrieval stress test. Mirror-to-regular mAP ratio for the 14 frozen models (foreground-only cutouts). This figure reports a retrieval-level check, not the canonical definition of Axis 2, which remains mean mirror similarity under each model’s native retrieval score. Lower values indicate greater asymmetry awareness. Green: pattern-aware (

<0.90

); orange: moderate (

0.90

–

0.95

); red: laterality-agnostic (

>0.95

). The dashed line at

1.0

marks the point where mirrored queries perform as well as originals. Only four models fall below 0.90 (ResNet152, MiewID-MSv2, DINOv3-ViT Pooled, I-JEPA), while EVA-02 is mirror-preferred (ratio

>1.0

Rank	Model	Regular mAP	Mirror mAP	Ratio	Tier
1	ResNet152	0.215	0.184	0.857	Pattern-Aware
2	MiewID-MSv2	0.295	0.257	0.871	Pattern-Aware
3	DINOv3-ViT Pooled	0.278	0.243	0.874	Pattern-Aware
4	I-JEPA	0.238	0.209	0.879	Pattern-Aware
5	ResNet50	0.227	0.215	0.945	Moderate
6	meFeM-B	0.199	0.192	0.962	Lat.-Agnostic

Appendix

Table 23: Auxiliary mirrored-query retrieval ratio (mirror / regular queries) on foreground-only images. Numerical data behind Figure 10 . Axis 2 itself is mean mirror similarity under the model’s native retrieval score. EVA-02 is the only model where mirror queries retrieve better than originals (ratio

>1.0

), consistent with its near-perfect mirror similarity (

0.997

Panel of Figure 11 — Figure 11: Two positive-danger-margin cases according to MegaDescriptor-L embeddings. Each row: original foreground crop (left), horizontal mirror (centre), nearest wrong-identity match (right). In both cases the mirrored image is less similar to the original than a different individual ( $\text{danger\_margin}>0$ ), confirming that horizontal flips can corrupt identity for laterality-aware models. Both images have below-average solidity ( $0.65$ and $0.54$ vs. dataset mean $0.69$ ), suggesting poorer segmentation increases vulnerability ( Table 24 ).

Identity	Images	DM $>0$	Mean DM	Max DM
Overa	62	1 (1.6%)	$-$ 0.337	$+$ 0.072
Medrosa	170	1 (0.6%)	$-$ 0.378	$+$ 0.053
Katniss	63	0	$-$ 0.291	$-$ 0.000
Apeiara	20	0	$-$ 0.370	$-$ 0.013
Ousado	179	0	$-$ 0.356	$-$ 0.034
All 31 identities		2 (0.1%)	$-$ 0.366	$+$ 0.072

Appendix

Table 24: Danger margin scan across all 1,895 training images (MegaDescriptor). Only 2 images (0.1%) have positive danger margin, both from low-contrast identities. Identities sorted by max danger margin; top 5 shown. Aggregate: mean

\text{danger\_margin}=-0.366

, median

-0.369

Model	Raw mirror	Mirror rank pct	Mirror z	Mean danger_margin
ArcFace O0 (MSv3)	0.855 $\pm$ 0.004	1.000 $\pm$ 0.000	0.798 $\pm$ 0.011	-0.412 $\pm$ 0.006
ArcFace O1 (MSv3)	0.847 $\pm$ 0.002	0.667 $\pm$ 0.000	0.748 $\pm$ 0.002	-0.391 $\pm$ 0.004
Lorentz O0 (MSv3)	0.290 $\pm$ 0.002	0.333 $\pm$ 0.471	-0.260 $\pm$ 1.192	-0.146 $\pm$ 0.003
Lorentz O1 (MSv2)	0.291 $\pm$ 0.002	0.333 $\pm$ 0.236	-0.359 $\pm$ 0.615	-0.118 $\pm$ 0.001
Lorentz O1 (MSv3)	0.296 $\pm$ 0.008	0.833 $\pm$ 0.236	0.619 $\pm$ 0.788	-0.149 $\pm$ 0.008
Triplet T0 (MSv3)	0.741 $\pm$ 0.002	0.333 $\pm$ 0.000	0.124 $\pm$ 0.018	-0.340 $\pm$ 0.004

Appendix

Table 25: Axis 2 summary from the active v5 TEST bundle. Cross-space claims use mirror_rank_pct_mean (lower is better). Raw mirror similarity and z-normalised values are shown for context only and do not override the rank-primary ordering.

Model	Input	Preprocess	Global descriptor	Dim
DINOv3-ViT (CLS)	224	HF processor	CLS token	384
DINOv3-ViT (pooled)	224	HF processor	Mean pool patch tokens	384
DINOv3-ConvNeXt (CLS)	224	HF processor	CLS token	768
DINOv3-ConvNeXt (pooled)	224	HF processor	Mean pool patch tokens	768
MiewID-MSv2	440	Resize + ImageNet norm	Model embedding (GeM + BN)	2152
MiewID-MSv3	440	Resize + ImageNet norm	Model embedding (GeM + BN)	2152

Appendix

Table 26: Frozen embedding extraction details for the 15 deep backbones from the frozen benchmark (main Table 1; HotSpotter+VLAD excluded as classical CV). Input sizes and global descriptor choices follow each model’s canonical evaluation wrapper. “HF processor” denotes the HuggingFace image processor; “timm config” denotes timm’s resolved preprocessing configuration.

Animal Re-Identification

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

See all 33 papers

Aug 18, 2026

Visual-Prompt Guided Wildlife Instance-Level Recognition

Fine-grained wildlife re-identification remains a challenging area in research. Current state-of-the-art approaches apply a detection and re-identification pipeline. We propose a one-stage end-to-end detection and re-identification model that performs identity searching within the latent space. We adopt DINOv2 for robust spatial geometry and MegaDescriptor for wildlife re-identification. We enhance latent queries with prompt re-identification features. A detection decoder queries the scene latent space to establish object boundaries around the target identity. Preliminary findings reflect a competitive mean average precision score of 30.584% compared to the state-of-the-art two stage approach of 44.89%. Qualitative results depict effective bounding and identification of animal identities.

Mufhumudzi Muthivhi, Jiahao Huo, Terence van Zyl +1

University of Johannesburg, South Africa · Linköping University, Sweden

Animal Re-Identification Object Detection

Ask PDF

Mar 4, 2026

Degradation-based augmented training for robust individual animal re-identification

Wildlife re-identification aims to recognise individual animals by matching query images to a database of previously identified individuals, based on their fine-scale unique morphological characteristics. Current state-of-the-art models for multispecies re- identification are based on deep metric learning representing individual identities by fea- ture vectors in an embedding space, the similarity of which forms the basis for a fast automated identity retrieval. Yet very often, the discriminative information of individual wild animals gets significantly reduced due to the presence of several degradation factors in images, leading to reduced retrieval performance and limiting the downstream eco- logical studies. Here, starting by showing that the extent of this performance reduction greatly varies depending on the animal species (18 wild animal datasets), we introduce an augmented training framework for deep feature extractors, where we apply artificial but diverse degradations in images in the training set. We show that applying this augmented training only to a subset of individuals, leads to an overall increased re-identification performance, under the same type of degradations, even for individuals not seen during training. The introduction of diverse degradations during training leads to a gain of up to 8.5% Rank-1 accuracy to a dataset of real-world degraded animal images, selected using human re-ID expert annotations provided here for the first time. Our work is the first to systematically study image degradation in wildlife re-identification, while introducing all the necessary benchmarks, publicly available code and data, enabling further research on this topic.

Thanos Polychronou, Lukáš Adam, Viktor Penchev +1

School of Mathematical Sciences, Queen Mary University of London, United Kingdom · University of West Bohemia in Pilsen, FEE, RICE, Pilsen, Czechia

Animal Re-Identification Image Corruption Robustness Data Augmentation

Ask PDF

Sep 5, 2026

What Does Animal Re-Identification Learn? Linear Biological Concepts and Their Origins in Visual Representations

Conservation increasingly relies on camera traps that collect more wildlife imagery than experts can manually analyze, making animal re-identification (Re-ID) essential for monitoring individuals and populations. Yet understanding which cues drive model decisions is challenging for ViT-based Re-ID models, whose metric-learning objectives provide no explicit supervision for biological concepts. We ask whether such models nonetheless organize their representations along biologically meaningful axes. Using a DINOv3 backbone fine-tuned for Western lowland gorilla Re-ID with triplet-margin loss, we find that sex and age emerge as linear directions that generalize to held-out individuals, reaching up to 0.91 AUROC and being recoverable from a single image per individual. Activation steering further shows that the sex direction is causally used by the model, flipping a significant fraction of predictions to the opposite sex. Comparing off-the-shelf and fine-tuned backbones shows that Re-ID training does not create these concepts, but relocates them across the network. Finally, data attribution reveals that the representation we find reflects a graded biological axis, is redundantly encoded across the population and shaped by visually ambiguous individuals. Together, these findings show how interpretability can uncover both the biological structure and failure modes of Re-ID representations, providing a step toward auditable computer vision for wildlife monitoring.

Robert Nolting, Alexandra Schild, Moritz Weckbecker +2

Hasso-Plattner Institute, Germany · University of Potsdam, Germany · Department of Artificial Intelligence, Fraunhofer Heinrich Hertz Institute, Germany

Animal Re-Identification Visual Representation Learning Concept-Based Explanations Wildlife Monitoring

Ask PDF

Abstract

Animal Re-Identification

Explore similar work

Visual-Prompt Guided Wildlife Instance-Level Recognition

Degradation-based augmented training for robust individual animal re-identification

What Does Animal Re-Identification Learn? Linear Biological Concepts and Their Origins in Visual Representations