OpenStripe supports open-set identification and spatial stripe profiling in wild Amur tigers

AI summary

Overview

Research area: Computer vision for wildlife biometrics, combining a large-language-model multi-agent pipeline with ecological population monitoring (animal re-identification, spatial phenotyping).

Technical level: Advanced. The paper mixes LLM-orchestrated image reconstruction, topological feature encoding, nine clustering pipelines, consensus clustering, and spatial statistics (Mantel tests, quantile regression, bootstrap confidence intervals).

Scope (one sentence): The authors introduce OpenStripe, a training-free multi-agent framework that rebuilds canonical 2D tiger torso stripe patterns, encodes four interpretable local stripe structures, and uses them for open-set individual clustering and geographic profiling of 28 wild Amur tigers.

What This Paper Is About

Camera traps generate huge numbers of images of endangered animals, but turning those images into individual identities is hard: animals bend and twist, vegetation blocks the view, lighting changes, and — critically — new animals keep appearing that were never in any training set. Existing re-identification models assume a closed world where every test individual was seen during training, so they break down in real field monitoring. OpenStripe addresses this by having a team of LLM-driven agents reconstruct a clean, flattened 2D stripe map of each tiger's torso and label interpretable stripe structures, then cluster images into identities without ever being told how many tigers exist.

Key Contributions

1. A multi-agent reconstruction-and-annotation framework. OpenStripe coordinates four agents (foreground extraction, source-aligned reconstruction, structural annotation, final review) through a shared blackboard architecture, all using GPT Image 2.5 via ChatGPT, to produce monochrome 2D torso stripe reconstructions annotated with four local structures: breakpoints (B), bifurcations (F), irregular structures (I), and round structures (C).

2. An open-set clustering method that needs no predefined class list or cluster count. Combining all four annotation types with source-balanced consensus clustering reached 0.90 matched accuracy and 0.85 adjusted Rand index (ARI) on 100 camera-trap records from 28 wild Amur tigers, without being given the true number of clusters.

3. A large benchmarking effort against existing approaches. The authors evaluated 49 baseline configurations (41 single-feature, 8 fusion) built from MegaDescriptor, MiewID, DINOv2, and other pretrained encoders under the same clustering backend, plus nine clustering pipelines and 17 input representations in ablation.

4. A link between stripe topology and geography. The extracted features support spatial phenotypic profiling, showing that stripe similarity declines with geographic distance and identifying six topological traits that carry geographic signal.

Main Findings

• Reconstruction quality: Foreground segmentation reached 91.20% mean intersection over union (mIoU), beating GrabCut by 25.80 percentage points and Mask R-CNN by 7.70 percentage points (paired Wilcoxon signed-rank tests, p < 0.001).

• Pose alignment quality: Torso alignment achieved 95.40% PCK@0.1 and 3.10% normalized mean error (NME), outperforming thin-plate splines (ΔPCK@0.1 = 23.60 pp, ΔNME = −5.40 pp) and spatial transformer networks (ΔPCK@0.1 = 9.20 pp, ΔNME = −2.10 pp), all p < 0.001.

• Clustering performance: OpenStripe with consensus clustering reached 0.90 accuracy, 0.85 ARI, 0.97 normalized mutual information, and 0.92 B-cubed F1. The circular dendrogram used a discovery threshold of 0.61 and produced 10 misassigned records.

• Baselines fell far short: The strongest of 49 baselines was "Animal-SNF" — similarity network fusion combining MegaDescriptor-L, MiewID, and DINOv2-B/224 — at 0.43 accuracy and 0.14 ARI. OpenStripe's edge was 47.00 pp in accuracy and 70.70 pp in ARI.

• More annotation types is better: Across all nine clustering pipelines, the full B/F/I/C combination performed best, with a positive trend as annotation types were added (β_accuracy = 0.20, β_ARI = 0.21, paired t-test, p < 0.001).

• Bifurcations matter most: Mean accuracy gains were 16.57 pp for B, 41.29 pp for F, 7.14 pp for I, and 24.00 pp for C, corresponding to shares of 18.62%, 46.39%, 8.03%, and 26.97% of the summed marginal gains. Adding appearance descriptors gave a further Δaccuracy = 0.02 and ΔARI = 0.03.

• Stripe similarity decays with distance: Spearman Mantel correlations with geographic distance were ρ = −0.36 for stripe similarity and ρ = −0.41 for annotation similarity (p < 0.001), and each of the four annotation types individually showed a significant negative correlation (p < 0.050).

• Topology beats pixel overlap: Morphological similarity fit distance better than Pixel IoU (R² = 0.18 versus R² = 0.05), with significant Mantel signal in the 0.00–18.16 km class (p = 0.010 versus p = 0.398) and the 56.38–110.00 km class (p = 0.008 versus p = 0.204).

• Six informative traits: Of 18 extracted topological traits, six exceeded the SNR = 1.00 threshold — fractal dimension, relative inter-stripe spacing, dorsoventral density gradient, stripe orientation (sin 2θ), isolated spot fraction, and stripe width IQR/median ratio — and all six showed similarity declining with distance.

Methodology in Plain English

The team started with 100 camera-trap photos of 28 individual Amur tigers taken in 2023 in Northeast China Tiger and Leopard National Park and adjacent habitat along the China–Russia border, with 1–5 records per individual. Because raw photos are distorted by pose and clutter, they built a chain of four LLM agents that work like a relay with a shared notepad: one cuts the tiger out of the background, one warps it onto a standard flat torso view, one labels stripe structures, and one checks the result. The output is a clean black-and-white stripe drawing rather than a photograph.

Instead of treating the stripe pattern as a whole image to compare pixel by pixel, the framework turns it into a graph of stripe shapes: where a stripe ends (B), where it splits into three (F), where something messy happens (I), and where a closed loop or isolated dot appears (C). These are defined by how many neighbours a stripe point has — degree 1, 3, at least 4, or 0/first Betti number 1. Two tigers are compared by matching where their labelled structures sit on the standardized torso.

For grouping images into individuals, the authors tried nine clustering algorithms (Leiden, k-reciprocal HDBSCAN, K-means, FINCH, HDBSCAN, agglomerative hierarchical clustering, two DINOv2-augmented variants, and consensus clustering) and 17 input representations. Their consensus method pools 27 representation–method groups, keeps the top 50% of partitions in each group, and averages co-assignment matrices. Cluster count is chosen automatically from K = 2 to 60 by an internal objective balancing silhouette score and coverage, with penalties for singleton and oversized clusters. Identities were verified by expert correction of the initial automated clustering.

For the geography analysis, each tiger's location was summarized as a spherical mean and pairwise distances were computed with the Haversine formula, excluding two pairs beyond 110 km to leave 376 valid pairs. Associations were tested with Spearman Mantel tests using 9,999 permutations and 1,999 bootstrap resamples, and 10 km bins were fitted with ordinary least squares.

Why This Matters

Impact on research. The paper argues that closed-set assumptions are a fundamental mismatch for field ecology, where births, dispersal, and migration constantly add unseen individuals. OpenStripe shows a training-free, interpretable alternative that also produces data usable for downstream biology — the authors argue the four BFIC structure types form an approximately complete topological basis that could extend to zebra stripes, leopard rosettes, cheetah spots, and giraffe patches, all of which arise from Turing reaction–diffusion patterning. The work also provides indirect evidence that stripe topology carries kinship signal, a starting point for genetic studies. The authors note that the mean sample size per individual in their dataset is only 3.57, and the median across 34 other wild-animal identification datasets is just 8.81 — a data regime where pretrained deep models struggle.

Real-world applications:

• Non-invasive population censuses and abundance estimates for endangered carnivores, where new individuals must be recognized without retraining.
• Long-term monitoring of dispersal across the China–Russia border, where the Amur tiger population is actively expanding into new habitat.
• Transferable individual identification for other patterned species (zebras, cheetahs, leopards, snow leopards, giraffes, whale sharks) using the same topological vocabulary.
• Spatial phenotypic profiling to flag likely related individuals and prioritize where to invest in genetic sampling such as fecal DNA.

Industry relevance. The blackboard multi-agent pattern with a vision model as the rendering engine is a template for any domain needing structured, interpretable outputs from messy imagery — inspection, medical or veterinary imaging, and other biometrics. The demonstration that a training-free agent pipeline beats specialized wildlife encoders on a small dataset is directly relevant to conservation-tech organizations, camera-trap software vendors, and national park monitoring programs that cannot afford large labelled training sets.

Future Directions

• Two-flank problem. Results rest on single-flank records; B/F/I/C markers on the unobserved opposite flank could reduce accuracy. The authors plan to address this after establishing links between flanks.
• Missing sex information. Without sex data, the authors cannot separate the bias that long-distance dispersing males introduce into geographic statistics. They propose combining sex determination with fecal DNA identification.
• Genetic validation. The kinship interpretation of stripe topology is indirect; the topological traits extracted by OpenStripe remain to be tested against actual genetic relatedness.
• Generalization and scale. The evaluation covers 28 individuals from one population; scaling to other species and larger cohorts, and integrating structured foundation models or world models into the agent framework, are stated as open directions.

Target Audience

Conservation biologists and wildlife ecologists running camera-trap programs; computer vision researchers working on animal re-identification and open-set recognition; machine learning engineers interested in multi-agent LLM pipelines for structured visual tasks; and park managers or conservation-tech practitioners who need identification tools that work with small, unlabeled field datasets. Readers without a background in clustering metrics or spatial statistics will find the conceptual framing accessible, but the evaluation sections assume familiarity with ARI, NMI, B-cubed F1, and Mantel tests.

Authors’ abstract

Non-invasive camera-trap monitoring of endangered wildlife requires open-set individual identification---grouping field images into identities without predefined classes or a known population size---yet pose deformation and cluttered backgrounds complicate image comparison. Here, we present OpenStripe, a multi-agent framework that reconstructs canonical two-dimensional (2D) tiger torso patterns and encodes four interpretable local stripe structures: breakpoints, bifurcations, irregular structures, and round structures. We evaluated OpenStripe on 100 camera-trap records from 28 wild Amur tigers (Panthera tigris altaica), representing nearly one-third of the documented population in the Northeast China Tiger and Leopard National Park. Combining all four complementary structures enabled automatic identity grouping with 0.90 matched accuracy and a 0.85 adjusted Rand index (ARI), outperforming 49 baseline configurations under the same backend (up to 0.43 accuracy and 0.14 ARI). In addition, results from OpenStripe indicate that stripe similarity in Amur tigers declines with geographic distance and suggest that traits such as fractal dimension and inter-stripe spacing are associated with geographic proximity. OpenStripe thus provides an interpretable representation that connects automated individual grouping with spatial phenotypic profiling in wild populations.