cs.CVSep 5, 2026

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

Authors: Robert Nolting, Alexandra Schild, Moritz Weckbecker, Maximilian Schall, Gerard de Melo

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

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

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.

Figures & tables

Figure 1: Feature attribution for a female test image, produced by DualXDA. The middle image shows a female gorilla that was correctly predicted by our activation probe, together with an LRP heatmap showing which parts of the image were important for the classification. The right side shows the training images that pulled the prediction most towards female, the left side towards male. The heatmaps show that relevance concentrates on facial and anatomical regions but also on the nursing infant, suggesting a potential shortcut.
Figure 2: Layer-wise linear separability of biological concepts in the activations of a ViT fine-tuned for gorilla Re-ID, measured in AUROC.
ClassSexAgeTrain Ind.Test Ind.Images
SilverbackMA1066,477
BlackbackMA562,962
Adult FemaleFA31923,132
Adolescent & Juvenile/n-A21810,066
Infant/n-A732,595
Total  753345,232
Table 1: Dataset composition by age-sex class. Sex is denoted as M/F (male/female) and age as A/n-A (adult/non-adult). Silverbacks are sexually mature males with silver hair; Blackbacks are maturing males; Adolescents & Juveniles are non-adults without reliable sex traits; and Infants are young dependent individuals.
Figure 3: Concept separability with the full training set versus one image per individual. Shading indicates one standard deviation across 32 random single-image samples.
Figure 4: Cross-model transfer between probes trained on fine-tuned activations (ft →\rightarrow pre) and pre-trained-only activations (pre →\rightarrow ft). In-domain probes show the normal probes’ AUROCs.
Figure 5: Percentage of steered queries flipping their prediction towards an individual of the opposite sex compared to the baseline. Orange shows the embedding degradation as percentage of steered queries leaving the defined embedding manifold. The layer-wise plot shows that steering only becomes meaningful at middle layers. Increasing steering dose ( α\alpha ) produces more steering effect but also higher degradation.
Figure 6: The ID-CV probe’s mean logit for each held-out individual. Individuals form a graded axis, with silverbacks and adult females occupying opposite extremes and blackbacks concentrated between them.
Figure 7: Enrichment of silverbacks and blackbacks among the constituting examples of the sex probe. Enrichment is defined as the ratio (share in the SVs/k90 / share in the training pool). Blackbacks become increasingly over-represented in deeper layers.
Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 8: Percentage of steered queries that switch to an individual of the opposite age. Orange denotes off-manifold degradation. Steering becomes effective in the middle layers and increases with α\alpha , alongside greater degradation.
Figure 9: ID-CV probe’s mean age-probe logit for each held-out individual. The same graded axis as with the sex-probe emerges, with the distinctive silverbacks and infants making up the extremes.
Figure 10: DualXDA heatmap for the ’age’ probe, evaluated on an infant gorilla. The same Re-ID relevant regions are highlighted.
Figure 11: Layer-wise AUROCs of LR sex probes on the jaguar and polar bear datasets, with one identity or one zoo held out. “Per identity” denotes aggregating logits over all images of an identity and evaluating the resulting identity-level predictions.
DatasetImagesIDsMaleFemale
Jaguar [ 26 ]1.86130822
PolarBearVidID [ 31 ]138.3631349
Appendix
Table 2: Statistics of the additional Re-ID datasets. A dash indicates unavailable information.
Figure 12: Cumulative sum of attributions against x% of training images included at layer 7. Dashed lines indicate regularization C (0.001, 0.1 and 10), colors indicate probe type.
Figure 13: Layer-wise linear separability of the sex and age concepts for the three probe types
Figure 14: Comparison between an LR probe fitted on pre-trained-only activations vs on fine-tuned activations.
Figure 15: Cosine similarity between concept directions learned on pretrained and Re-ID-fine-tuned activations.
Figure 16: Steering results for LR sex and age probes at layer 9 across different values of KK and steering strength α\alpha . All results are reported as percentages.
Figure 17: Number of images in the SV vs k90 sets for sex and age concepts
Figure 18: Percentage of each ID’s total images selected to be support vectors (layer 9).
More in this topic

Animal Re-Identification

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

See all 33 papers

Explore similar work

CardsList

  1. Visual-Prompt Guided Wildlife Instance-Level Recognition

    Aug 18, 2026Mufhumudzi Muthivhi, Jiahao Huo, Terence van Zyl +1Animal Re-IdentificationObject Detection

    University of Johannesburg, South Africa · Linköping University, Sweden

  2. Spotted: Location-informed Reidentification of Hyenas and Leopards in Camera Trap Surveys

    Jul 1, 2026Halil Sina Kelebek, Julia Hindel, Kobus Hoffman +9Animal Re-IdentificationHuman-in-the-Loop Annotation

    Oxford Robotics Institute, University of Oxford, Oxford, UK · Department of Computer Science, University of Freiburg, Germany · Bubye Valley Conservancy, Zimbabwe +3

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

    Apr 6, 2026Antonio Rueda-Toicen, Abigail Allen Martin, Daniil Morozov +5Animal Re-IdentificationShortcut Learning

    Hasso Plattner Institute · NVIDIA · Jaguar ID Project +5