PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Yichun Peng, Yansong Li, Jin Hou, Qi Song, Xi Guo, Ning Zhang, Degang Wei, Huang Yang, Yu Guan, Jianping Ge, Lei Bao, Tianming Wang

arXiv:2610.32848v1cs.CV

TL;DR

Automated wildlife identification methods often assume test-time individuals were included in training, although reproduction and migration introduce new individuals, and field images are complicated by coat deformation, vegetation occlusion, and illumination changes. OpenStripe is a multi-agent system that uses the four annotation types B, F, I, and C to support individual clustering. OpenStripe source-balanced consensus clustering (Con.) achieved an accuracy of 0.90 and an ARI of 0.85, outperforming Animal-SNF (accuracy = 0.43, ARI = 0.14; ∆accuracy = 0.47, ∆ARI = 0.71, paired t-test, p < 0.001).

  • Problem

    Automated wildlife identification methods often assume test-time individuals were included in training, although reproduction and migration introduce new individuals, and field images are complicated by coat deformation, vegetation occlusion, and illumination changes.

    Source: Introduction · p. 1

  • Method

    OpenStripe is a multi-agent system that uses the four annotation types B, F, I, and C to support individual clustering.

    Source: Introduction · p. 1, OpenStripe supports open-set… · p. 2

  • Results

    OpenStripe source-balanced consensus clustering (Con.) achieved an accuracy of 0.90 and an ARI of 0.85, outperforming Animal-SNF (accuracy = 0.43, ARI = 0.14; ∆accuracy = 0.47, ∆ARI = 0.71, paired t-test, p < 0.001).

    Source: OpenStripe supports open-set… · p. 2

  • Takeaways & Limitations

    Stripe features and individual clustering results provide a basis for exploring geographic associations and potential stripe-phenotype inheritance in wild Amur tigers.

    Source: Introduction · p. 2

    The experiments used single-flank records from 28 Amur tigers, and markers on the unobserved opposite flank may reduce identification accuracy.

    Source: Future directions · p. 3

Abstract

from arXiv · show

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.

Introduction

OpenStripe addresses open-set wildlife identification and field-image challenges with a multi-agent approach to tiger stripe patterns. It reports strong Amur tiger clustering, accurate reconstruction and feature annotation, and geographic associations in stripe patterns.

Introduction

  • The authors developed OpenStripe as a multi-agent system for clustering Amur tigers in an open-set setting, where new individuals may enter through reproduction or migration.Existing automated methods are described as assuming test-time individuals were included in training.Source: Introduction · p. 1
  • In China Tiger and Leopard National Park, clustering achieved 90.00% accuracy and an adjusted Rand index (ARI) of 0.85, improving over the best baseline by 47.00 and 70.70 pp, respectively.Source: Introduction · p. 2
  • Stripe similarity declined with geographic distance, while fractal dimension, inter-stripe spacing, and dorsoventral density gradients showed spatial associations.The authors identify these patterns as a basis for investigating spatial structure and potential genetic relatedness.Source: Introduction · p. 2

OpenStripe accurately reconstructs Amur tiger stripe patterns and annotates key features

  • OpenStripe’s foreground segmentation reached 91.20% mIoU, outperforming GrabCut by 25.80 pp and Mask R-CNN by 7.70 pp.Both comparisons had p < 0.001.Source: OpenStripe accurately reconstructs… · p. 2
  • OpenStripe achieved 95.40% PCK@0.1 and 3.10% NME, outperforming both TPS and STN on these reconstruction measures.The reported differences from TPS and STN were significant at p < 0.001.Source: OpenStripe accurately reconstructs… · p. 2
  • OpenStripe annotates breakpoints, Y-shaped bifurcations, irregular structures, and round structures as features for clustering.The paper frames open-set identification as important for population monitoring and notes that coat deformation, vegetation occlusion, and illumination changes complicate field-image feature extraction.Source: OpenStripe accurately reconstructs… · p. 2

OpenStripe supports open-set individual clustering of Amur tigers

Across nine clustering pipelines, combining all four stripe annotation types performed best, and more annotation types were associated with better clustering. OpenStripe’s source-balanced consensus achieved 0.90 accuracy and 0.85 ARI, outperforming the strongest of 49 baseline configurations.

OpenStripe supports open-set individual clustering of Amur tigers

  • The complete BFIC combination achieved the best performance across nine pipelines, and more annotation types yielded better accuracy and ARI (βaccuracy = 0.20, βARI = 0.21; p < 0.001).The ablation tested unannotated full-body stripe images, unannotated stripes, and all 15 nonempty combinations of the four types.Source: OpenStripe supports open-set… · p. 2
  • 0.90 accuracy and 0.85 ARI were achieved by source-balanced consensus clustering, versus 0.43 accuracy and 0.14 ARI for the strongest baseline, Animal-SNF.OpenStripe significantly outperformed Animal-SNF (∆accuracy = 0.47, ∆ARI = 0.71, p < 0.001) among 49 configurations; consensus clustering was not given the true number of clusters.Source: OpenStripe supports open-set… · p. 2
  • Adding appearance descriptors further improved clustering performance, with ∆accuracy = 0.02 and ∆ARI = 0.03.The four annotation types yielded mean accuracy gains of 16.57, 41.29, 7.14, and 24.00 pp for B, F, I, and C, respectively.Source: OpenStripe supports open-set… · p. 2

OpenStripe supports exploration of geographic associations in Amur tiger stripe patterns

OpenStripe-derived stripe features were associated with geographic distance in the 28 Amur tigers, and six topological traits were identified as potential priorities for inheritance research. The authors also report that holistic visual encoders performed poorly in this open-set setting, whereas OpenStripe uses cross-modal reasoning and is training-free.

OpenStripe supports exploration of geographic associations in Amur tiger stripe patterns

  • Stripe similarity and annotation similarity were negatively correlated with geographic distance (Spearman Mantel ρ = −0.36 and −0.41, respectively; p < 0.001).Similarity in each of the four annotation types was also negatively correlated with distance (ρB,F,I,C < 0, pB,F,I,C < 0.050).Source: OpenStripe supports exploration of… · p. 2
  • Morphology characterized similarity better than Pixel IoU (R2 = 0.18 vs 0.05), and six traits exceeded the SNR > 1.00 threshold: fractal dimension, relative inter-stripe spacing, dorsoventral density gradient, stripe orientation, isolated spot fraction, and stripe-width IQR/median.Similarities in all six traits were negatively associated with geographic distance.Source: OpenStripe supports exploration of… · p. 2
  • The authors describe the stripe-feature and clustering results as a way to explore phenotype–geographic-distance associations and provide indirect evidence concerning stripe-phenotype inheritance.They identify the six traits as potential priorities for future studies of phenotypic inheritance in Amur tigers and other striped mammals.Source: OpenStripe supports exploration of… · p. 2

OpenStripe has advantages in open-set identification

  • The authors report that pretrained visual encoders performed poorly amid pose deformation, occlusion, and background clutter, while OpenStripe’s cross-modal reasoning addresses these difficulties.They note a mean of 3.57 samples per individual in this study and describe multi-agent systems as training-free.Source: OpenStripe has advantages in open-set… · p. 3

Stability and completeness of the four annotation types

OpenStripe represents nonstandard stripe structures with four annotation types, which the authors describe as approximately complete for topological changes from ordinary stripes. The discussion connects the statistical properties of these structures with coat-pattern dynamics and geographic associations.

Stability and completeness of the four annotation types

  • OpenStripe defines B as breakpoints, F as degree-3 bifurcations, I as degree-4-or-higher or high-curvature nonstandard contours, and C as degree-0 structures or isolated loops.The authors state that these four types approximately cover topological changes that deviate from ordinary degree-2 stripes.Source: Stability and completeness of the four… · p. 3
  • The authors propose BFIC as a potentially universal topological basis for identification across mammals with stripes, rosettes, spots, or polygonal patches.They describe these coat patterns as solutions of two-component Turing reaction–diffusion PDEs under different parameter settings.Source: Stability and completeness of the four… · p. 3

Stripe topology reflects potential kinship structure

  • The discussion links geographic associations in Amur tiger stripe patterns with potential kinship associations.The authors state that local B/F/I/C positions are random, whereas their statistical properties are regulated by dynamical parameters.Source: Stripe topology reflects potential… · p. 3
  • Morphology fits the geographic-distance relationship better than Pixel IoU, and the four stripe-feature types are described as indicators of geographic associations.The authors suggest that future research on Amur tiger stripe genetics can build on extracted topological features.Source: Stripe topology reflects potential… · p. 3

Future directions

The study identifies limitations of its current experiments and outlines follow-up work on flank coverage and sex-related geographic statistics. It reports that no new physical materials or reagents were generated.

Future directions

  • Experiments use single-flank records from 28 Amur tigers, so markers on the unobserved flank may reduce identification accuracy.Future work will investigate this after establishing links between the two flanks.Source: Future directions · p. 3
  • Without sex information, geographic statistics cannot separate potential bias from long-distance-dispersing males; future work proposes combining sex determination with fecal DNA identification.The stated aim is to investigate genetic traits in greater depth.Source: Future directions · p. 3

Materials availability

Figure legends

The resource listings identify deposited OpenStripe data and the software and algorithms used for nonspatial analysis.

REAGENT OR RESOURCE SOURCE IDENTIFIER

Experimental model and study participant details

The study analyzed camera-trap records from identified wild Amur tigers in and around the Northeast China Tiger and Leopard National Park. Individuals had between one and five records, and ground-truth identities were expert-corrected.

Data details

  • 100 noninvasive camera-trap records from 28 wild Amur tiger individuals were analyzed.Source: Data details · p. 6
  • Records were collected in 2023 across the park and adjacent habitats along the China–Russia border, with individuals designated T01–T28 and represented by 1–5 records each.Source: Data details · p. 6
  • Ground-truth individual identities were obtained by expert correction of the initial OpenStripe + Con. clustering results.Source: Data details · p. 6

Reconstruction and annotation

OpenStripe coordinates agents to reconstruct tiger stripe patterns and annotate local structures, then evaluates reconstruction and foreground segmentation. The clustering methods include a source-balanced consensus procedure with internal cluster-count selection.

Reconstruction and annotation

  • OpenStripe coordinates four agents for foreground extraction, source-aligned reconstruction correction, structural annotation, and final correction using GPT Image 2.5 via ChatGPT.It annotates breakpoints (B), bifurcations (F), irregular structures (I), and round structures (C).Source: Reconstruction and annotation · p. 6
  • Foreground segmentation is evaluated with mIoU against GrabCut and Mask R-CNN, while reconstruction is evaluated with PCK@0.1 and NME against TPS and STN.Higher mIoU and PCK and lower NME indicate better performance; NME normalizes keypoint error by the source-image diagonal.Source: Reconstruction and annotation · p. 6
  • Skeleton endpoints E may include crop truncations, whereas B marks intrinsic stripe terminations or interruptions verified against the paired photograph and full reconstruction.Source: Reconstruction and annotation · p. 6

Clustering and automatic cluster selection

  • The study compares nine clustering pipelines, including Leiden, K-means, FINCH, HDBSCAN, AHC, DINOv2-augmented approaches, and source-balanced consensus clustering.The listed pipelines also include k-reciprocal reranking with HDBSCAN.Source: Clustering and automatic cluster… · p. 6
  • Con. averages co-assignment matrices from selected partitions across 27 representation–method groups, then applies average-linkage AHC to the complementary consensus distance.It selects cluster count over K = 2–60 using an objective combining silhouette and nonnoise coverage with penalties for singleton and oversized clusters; invalid partitions are excluded.Source: Clustering and automatic cluster… · p. 6

Ablation and baseline comparisons

The section specifies how annotation-count and per-type ablations were evaluated, and defines the feature and distance inputs for spatial analysis.

Ablation and baseline comparisons

  • The ablation analysis compared equally weighted mean clustering scores across zero to four annotation types and measured each type’s marginal gain across seven combinations of the other types.The 49 baselines comprised 41 single-feature and eight fusion configurations.Source: Ablation and baseline comparisons · p. 6

Spatial inputs and distances

  • Tiger locations were summarized by spherical-mean geographic centers, and Haversine distances were calculated for 376 valid pairs after excluding two pairs over 110 km.Source: Spatial inputs and distances · p. 7
  • Unannotated torso features used HOG histograms and stripe-density vectors, while B/F/I/C structures were compared through Hungarian matching of normalized coordinates.The torso regions of interest measured 256 × 128 pixels.Source: Spatial inputs and distances · p. 7

Morphological measurements and similarity

Morphological similarity was derived from standardized torso traits and compared with pixel overlap, while clustering and spatial analyses used specified metrics and resampling procedures.

Morphological measurements and similarity

  • Eighteen normalized, dimensionless torso traits were extracted after binarization and denoising; eight core features define Mahalanobis Morphology distance, rescaled to similarity.For seven local-window features, the measurement protocol was selected blindly using the intraclass correlation coefficient alone.Source: Morphological measurements and… · p. 7

Quantification and statistical analysis

Related papers

This paper cites

Loading 2610.32848v1…