Skip to content

OpenAnimals: Revisiting Person Re-Identification for Animals Towards Better Generalization

  • Conference: ICCV 2025
  • arXiv: 2410.00204
  • Code: Submitted with the paper (OpenAnimals codebase)
  • Area: Human Understanding
  • Keywords: Animal Re-Identification, Person ReID Transfer, Open-Source Framework, Baseline Model, Cross-Species Generalization

TL;DR

This paper develops the OpenAnimals open-source framework, systematically revisiting the transferability of person re-identification methods to animal re-identification. It proposes ARBase, an animal-oriented strong baseline that substantially outperforms existing person ReID methods across multiple benchmarks.

Background & Motivation

Animal Re-Identification (Animal Re-ID) aims to identify individual animals within a given species, which is critical for wildlife conservation, population monitoring, and behavioral research. Although conceptually similar to Person Re-Identification (Person Re-ID), the two tasks differ fundamentally:

Species Diversity: Different species (hyenas, leopards, sea turtles, whale sharks) exhibit dramatically different visual appearances and behaviors.

Environmental Variability: Habitats range from savannas to oceans, introducing far greater variation than the relatively controlled urban settings of person ReID.

Pose Differences: Quadrupedal locomotion (hyenas/leopards) and aquatic movement (sea turtles/whale sharks) differ fundamentally from human bipedal walking.

Data Scarcity: Data collection and annotation in wild environments are difficult, resulting in far less available data than person datasets.

The central question is: Can the extensive techniques and methodologies accumulated in person re-identification be effectively transferred to animal re-identification? Existing research lacks a systematic analysis of this question.

Method

Overall Architecture

The work is organized into three parts:

  1. OpenAnimals Framework: A unified animal ReID platform built upon FastReID and WildLifeDatasets.
  2. Systematic Revisiting Experiments: Ablating key designs from person ReID methods (BoT, AGW, SBS, MGN) one by one on animal benchmarks.
  3. ARBase Model: A strong animal-oriented baseline constructed from insights gained in the revisiting experiments.

OpenAnimals Framework Design

Two core principles are followed:

  • Person ReID Compatibility: Inherits the core layers of FastReID, enabling seamless integration of state-of-the-art person ReID methods.
  • Multi-Species Support: Integrates the dataset organization strategy of WildLifeDatasets, supporting 30+ species within a unified framework.

The modular design encompasses five stages: Data, Backbone, Head, Loss, and Training & Testing.

ARBase Model Design

Drawing from key findings in the revisiting experiments, ARBase makes targeted, animal-oriented design choices across five modules:

Data Module: - Key Modification — Input Resolution: Person ReID uniformly uses portrait-aspect resolutions (e.g., \([256,128]\)) since humans are typically upright. Animals exhibit diverse poses; ARBase adopts a square resolution of \([384,384]\), a simple change with significant effect. - Only random horizontal flipping (\(p=0.5\)) is used; Random Erasing and AutoAug are removed.

Backbone Module: - ResNet-50 (ImageNet pre-trained) with last stride set to 1 for fine-grained features. - Instance-Batch Normalization (IBN) replaces standard BN: IN learns appearance-invariant features (adapting to diverse environments), while BN retains content information. - Multi-branch architecture: global branch + 2-part branch + 3-part branch (inspired by insights from MGN).

Head Module: Global Average Pooling + Linear + BNNeck (decoupling the feature spaces for triplet and cross-entropy losses).

Loss Module: - Triplet loss computed on features before BNNeck: \(L_{tp} = \frac{1}{N_b}\sum_{i=1}^{N_b}\text{max}(0, m + d_{pos}^i - d_{neg}^i)\) - Cross-entropy loss with label smoothing computed on features after BNNeck.

Training & Testing: Adam optimizer + Cosine Annealing learning rate schedule.

Experiments

Main Results: ARBase vs. Person ReID Methods

Method HyenaID R1/mAP LeopardID R1/mAP SeaTurtleID R1/mAP WhaleSharkID R1/mAP
BoT 58.64/34.96 54.92/27.65 84.01/41.92 52.54/20.86
AGW 56.36/32.72 54.10/28.67 85.17/46.18 50.76/21.11
SBS 51.82/30.56 51.23/26.54 84.01/44.63 47.46/18.84
MGN 55.91/31.08 53.69/28.21 86.05/46.67 50.25/21.47
ARBase 73.18/44.87 64.34/37.08 86.92/55.99 62.44/29.45

ARBase achieves Rank-1 improvements of 14.54% on HyenaID, 9.90% on WhaleSharkID, and 9.42% on LeopardID.

Key Findings from Revisiting Experiments

Technique Effective for Persons Generalization to Animals
Random Erasing ✓ ✗ (negative effect on 3/4 datasets; disrupts subtle individual details)
Label Smoothing ✓ ✓ (beneficial on ≥3 datasets)
Last Stride=1 ✓ ✓ (consistently beneficial)
BNNeck ✓ ✓ (consistently beneficial)
Non-local Attention ✓ ✗ (inconsistent effects)
Gen-mean Pooling ✓ ✗ (inconsistent effects)
Weighted Triplet ✓ ✗ (inconsistent effects)
Freeze Training ✓ ✗ (removal improves performance)
AutoAug ✓ ✗ (removal improves performance)
Cosine Annealing ✓ ✓ (consistently beneficial)
Multi-Branch (MGN) ✓ ✓ (multi-granularity features also effective for animals)

Ablation Study (Data & Backbone)

Configuration HyenaID R1/mAP WhaleSharkID R1/mAP
BoT [256,128] 58.64/34.96 52.54/20.86
BoT [384,384] 60.45/36.43 58.12/24.39
ARBase w/o IBN 69.09/43.58 61.93/29.28
ARBase w/o MB 71.36/42.87 61.42/27.78
ARBase (Full) 73.18/44.87 62.44/29.45

Simply adjusting the resolution to square raises BoT's Rank-1 on WhaleSharkID from 52.54% to 58.12% (+5.58%).

Ablation Study (Head, Loss, Training)

Configuration HyenaID R1/mAP WhaleSharkID R1/mAP
w/o BNNeck 64.55/39.23 44.42/22.29
w/o Label Smoothing 68.18/44.72 61.42/27.88
w/o Cosine Annealing 71.82/43.40 62.44/28.69
ARBase (Full) 73.18/44.87 62.44/29.45

BNNeck has a dramatic impact on WhaleSharkID (Rank-1 drops from 62.44% to 44.42%), confirming that decoupling the feature space is equally critical for animal ReID.

Highlights & Insights

  • The systematic revisiting experiments reveal that many techniques widely accepted as effective in person ReID—such as Random Erasing and AutoAug—do not transfer to animal ReID, as inter-individual differences in animals are more subtle and random erasing may destroy discriminative details.
  • The input resolution insight carries broad implications: the portrait-aspect resolutions long used in person ReID are entirely unsuitable for the diverse poses of animals, and switching to square resolution alone yields substantial gains.
  • The introduction of IBN elegantly addresses the high environmental variability inherent in animal ReID.
  • ARBase's design philosophy—simple yet targeted—achieves significant improvements without introducing complex modules.

Limitations & Future Work

  • Only four animal species are evaluated; generalization to a broader range of species (e.g., birds, insects) remains unexplored.
  • A fixed ResNet-50 backbone is used; stronger recent pre-trained models (e.g., DINOv2, CLIP) are not investigated.
  • The horizontal partitioning assumption of the multi-branch architecture may be inappropriate for certain animals (e.g., snakes, fish).
  • Spatiotemporal information from video sequences is not considered.
  • Dataset scales are relatively small, and in-depth analysis of overfitting and generalization is insufficient.
  • Person ReID: BoT (bag of tricks), AGW (non-local attention + GeM pooling), SBS (AutoAug + cosine annealing), MGN (multi-granularity network).
  • Animal ReID: HotSpotter (hand-crafted features), MegaDescriptor (multi-species pre-training), CLIP/DINOv2-based methods.
  • Open-Source Frameworks: FastReID (person), WildLifeDatasets (animal dataset management).

Rating

Dimension Score
Novelty ⭐⭐⭐
Effectiveness ⭐⭐⭐⭐⭐
Clarity ⭐⭐⭐⭐⭐
Practical Value ⭐⭐⭐⭐⭐
Overall 8.0/10