Viewpoint Estimation for Animal Re-Identification on Edge Devices

This repository contains the code and models to accompany my thesis at the Max Planck Institute for Intelligent Systems and University of Tübingen

• Species: Optimized for Equids (specifically Zebras)
• Input Dimensions: $224 \times 224$ bounding box crops
• Downstream Integrations: Embedded within RAPID and the Behaviors Inference Framework

---

See the demo of the viewpoint estimator on YouTube

Viewpoint.Estimation.for.Animal.Re-Identification.on.Edge.Devices.-.Demo.mp4

---

graph LR
    %% Style Definitions
    classDef data fill:#f4f7fa,stroke:#b0c4de,stroke-width:1.5px,color:#2c3e50,font-size:13px;
    classDef module fill:#fff5eb,stroke:#ffa07a,stroke-width:2px,stroke-dasharray: 0,color:#2c3e50,font-weight:bold,font-size:13px;

    %% Nodes
    A["Input Image<br>(224×224 Crop)"]:::data
    B["Custom YOLO<br>Pose Model"]:::module
    C["17 Keypoints with<br>confidence scores"]:::data
    D["Derived<br>Feature Set"]:::data
    E["Viewpoint<br>MLP"]:::module
    F["Regressed Angle<br>(φ)"]:::data
    G["Keypoint skeleton"]:::data

    %% Connections
    A --> B
    B --> C
    C --> D
    D --> E
    E --> F
    C -. Optional  .-> G

Loading

Repository Architecture

This repository is organized into two separate branches:

• master: Contains the full pipeline and model training scripts from my thesis. Use this branch to experiment, evaluate ablation configurations or train custom models.
• module: A lightweight, production-ready release containing only the core standalone estimator. Switch to this branch for direct, dependencies-minimized deployment and framework integration.

Binary model weights can be found in the release package for this thesis.

the demo/ directory includes sampled images for two zebras including 8 different viewpoints from the ZebraStereoID dataset.