PaperlayerView exact revision on arXiv

Source-linked AI summary

Cross-Species Animal Re-Identification with Semantic Consistency Learning

Shuoyi Chen, Yuejia Li, Mang Ye

arXiv:2609.09705v1cs.CV

TL;DR

Cross-species animal ReID is challenged by heterogeneous morphologies, appearances, and environments that fragment representations and limit generalization. SCL combines region-aware spectral normalization with dynamic cross-species neighborhood modeling to stabilize features and preserve shared semantic structure. Across 11 public datasets, it consistently improves over competitive state-of-the-art methods and supports generalization to unseen species and domains.

  • Problem

    Cross-species animal ReID must learn individual-discriminative representations despite heterogeneous morphologies, textures, and environments that produce species-specific feature fragmentation.

  • Method

    SCL combines FDSNorm, which suppresses environment-induced style variation while preserving structure, with CNM, which learns cross-species relational structures through dynamic neighborhoods.

  • Results

    SCL consistently improves over competitive state-of-the-art methods across 11 public animal ReID datasets and generalizes to unseen species and ecological domains.

  • Takeaways & Limitations

    SCL yields a unified, transferable embedding space by stabilizing feature statistics and capturing structural regularities shared across species.

  • Takeaways & Limitations

    FDSNorm uses an unconstrained variant without an explicit ordering constraint between foreground and background modulation parameters, leaving their relative strengths to the network.

Abstract

from arXiv · show

Generalizable animal Re-Identification (ReID) aims to recognize individual animals across species with diverse morphologies and ecological contexts. Unlike person ReID, where different domains share similar body structures, animal species often exhibit drastically different anatomical structures and visual patterns, making it difficult to establish shared visual correspondences. As a result, representations learned across species tend to form fragmented embedding spaces, which severely limits cross-species generalization. To address this challenge, we propose Semantic Consistency Learning (SCL), a framework designed to learn representations that remain stable across appearance variations while preserving semantic structures shared across species. SCL consists of two complementary components. Foreground-Background Decoupled Spectral Normalization (FDSNorm) stabilizes feature statistics by suppressing environment-induced style variations in a region-aware manner, while Cross-species Neighborhood Modeling (CNM) captures transferable relational structures across species through dynamic feature neighborhoods. Extensive experiments on 11 public animal ReID datasets demonstrate that SCL consistently outperforms state-of-the-art methods under multiple cross-species evaluation protocols and generalizes effectively to previously unseen species and ecological domains. Code is available at https://github.com/Kemalau/ECCV-26-SCL.

1 Introduction

Animal ReID faces distribution shifts from species, morphology, texture, habitat, and environmental context that exceed common viewpoint or illumination changes. SCL addresses this by stabilizing appearance statistics and modeling transferable cross-species relationships, with experiments showing consistent gains across 11 public datasets.

  • Motivation: Animal ReID must handle species, texture, morphology, habitat, and environmental shifts beyond viewpoint or illumination changes.These discrepancies make cross-species generalization a fundamental challenge.
  • Motivation: Single-species training limits model reusability because each new species requires additional data collection, identity annotation, and retraining.
  • Semantic Consistency Learning: SCL stabilizes appearance statistics while preserving semantic structures shared across species through region-aware frequency-domain normalization.FDSNorm preserves phase information and adaptively modulates amplitude spectra across foreground and background regions.
  • Semantic Consistency Learning: CNM learns transferable relational structures by dynamically constructing intra-species and inter-species neighborhoods through mutual neighbor discovery.The approach uses relational topology while preserving discriminative capability within each species.
  • Evaluation: SCL introduces cross-species evaluation protocols and consistently improves over competitive state-of-the-art methods on 11 public animal ReID datasets.

2 Related Work

Prior ReID research spans person and vehicle identification, domain generalization, and unified animal models. Existing approaches address style variation and large-scale animal recognition, but common metric-learning objectives do not explicitly target completely unseen species.

  • Person and vehicle ReID research includes convolutional, transformer-based, and vision-language architectures, but adapting standard frameworks requires separate models for each animal category.
  • Domain-generalized ReID methods address changes in scene, viewpoint, and illumination through normalization-based approaches and Mixture-of-Experts models.
  • Animal ReID foundation models, CLIP-based approaches, and community benchmarks support unified embedding learning and large-scale individual recognition across wildlife datasets.
  • Figure 2 presents Semantic Consistency Learning through Foreground–Background Decoupled Spectral Normalization and Cross-species Neighborhood Modeling.
  • Common metric-learning objectives do not explicitly design mechanisms for generalization to completely unseen species.

3 Method

SCL addresses fragmented cross-species representations with two complementary components: FDSNorm stabilizes appearance-sensitive features, while CNM models transferable relational structure across species.

  • Motivation: Cross-species animal ReID faces unstable feature statistics and species-dependent fragmentation because textures, morphologies, environments, and anatomical semantics differ substantially.These differences make alignment strategies and person-ReID domain-generalization assumptions unreliable across species.
  • Framework Overview: SCL combines region-aware spectral normalization with dynamic feature neighborhoods to stabilize representations and preserve transferable semantic structure.FDSNorm suppresses environment-induced style variation, while CNM aligns relational regularities within and across species.
  • FDSNorm: FDSNorm treats amplitude as style-related and phase as structure-related, then selectively normalizes foreground and background branches while preserving branch-wise phase information.The module uses spatial decoupling, branch-wise DFT, independent mixing strengths, and reconstruction with original phase.
  • FDSNorm: FDSNorm builds a soft foreground mask from CLS-to-patch cosine similarity and applies spectral normalization to patch tokens while keeping the CLS token unchanged.It is inserted at multiple Transformer depths to progressively address pixel-, structure-, and residual-level domain shifts.
  • Semantic Stabilization: Temporal semantic distillation uses an exponential-moving-average teacher to provide smoothed features that anchor neighborhood consistency learning.The teacher transfers long-term structural knowledge to the student without additional supervision.
  • CNM: CNM constructs intra-species and reciprocal cross-species neighborhoods from current teacher features and a memory queue to capture relational topology.Reciprocal filtering retains cross-species pairs only when neighborhood membership is mutual, producing more stable semantic neighborhoods.
  • CNM: CNM uses bounded margin-based attraction toward cross-species relational centers, providing connectivity without collapsing structural diversity.The hinge stops contributing once similarity reaches the margin, limiting excessive attraction between different identities.

4 Evaluation Protocol

The evaluation uses 11 public animal ReID datasets spanning diverse species, habitats, and visual conditions, with CMC Rank-1 and mAP as the primary metrics.

  • Datasets: 11 publicly available animal ReID datasets are used to evaluate cross-species generalization.
  • Datasets: The collection spans diverse habitats and species morphologies, creating substantial variation in appearance and environmental conditions.
  • Datasets: Wildlife71 contributes 71 species, with 67 seen species used for training under its official split protocol.
  • Metrics: Evaluation reports CMC at Rank-1 and mean Average Precision, including all valid gallery matches because most datasets lack camera annotations.

5 Experiments

Experiments evaluate SCL across unseen domains, protocols, ablations, and representation analyses. Results indicate improved cross-species generalization, structure-focused attention, and less species-isolated embedding structure.

  • Comparison with SOTA Methods: SCL achieves leading performance across diverse unseen species without target-domain adaptation, supporting species-agnostic cross-species generalization.The evaluation includes Protocol-1 results on 10 unseen domains and Protocol-2 results on Wildlife71.
  • Ablation Experiments: FDSNorm and CNM provide complementary contributions: FDSNorm stabilizes style statistics, while CNM promotes intra-species cohesion and cross-species semantic connectivity.The component ablation is conducted on ten wildlife datasets under Protocol-1.
  • Ablation Experiments: CNM memory significantly improves nearest-neighbor identity purity, whereas removing the memory queue causes a noticeable retrieval-purity drop.The memory queue supplies stable, long-term relational anchors for neighborhood modeling.
  • Visualization Analysis: SCL shifts activation toward semantically meaningful animal regions and reduces diffuse background attention across eight diverse species.The observed regions include torso contours, texture-rich parts, and limbs, with the trend remaining stable across large appearance and habitat variations.
  • Feature Distribution Analysis: CNM produces more mixed cross-species embedding distributions than the baseline while achieving better ReID performance.The visualization samples 100 instances from each of 22 species; the baseline forms isolated species-wise clusters.
  • Feature Distribution Analysis: SCL achieves higher inter-ID-to-intra-ID distance ratios than MegaDescriptor and MiewID on all ten unseen datasets.Higher ratios indicate tighter intra-identity compactness and clearer inter-identity separation.

6 Conclusion

The paper introduces SCL for cross-species generalization in animal ReID. FDSNorm stabilizes features while CNM learns cross-species relational structure, producing a transferable embedding space and motivating future open-set and ecological deployment extensions.

  • 6 Conclusion: SCL combines FDSNorm and CNM to suppress environmentally caused feature instability and capture structural regularities that generalize across species.The framework shifts learning toward relational semantic understanding and a unified embedding space.
  • 6 Conclusion: The proposed framework targets transferable animal ReID representations for applications including large-scale biodiversity monitoring in real-world ecosystems.Future directions include fully open-set ReID, ecological foundation models, and deployment on UAVs and camera traps.
Loading 2609.09705v1…