🐾 Wild Animal Re-Identification

You can find the full paper about this project here.

---

📌 Project Overview

This project aims to address the problem of wild animal re-identification (Re-ID) across various species such as lynx, sea turtles, and salamanders.

This system was developed to participate in the AnimalCLEF25 / LifeCLEF Wild Animal Re-Identification Challenge, part of the CVPR 2025 - FGVC10 Workshop.

To identify individuals accurately, the system leverages multiple global and local descriptors, combining them through Feature Fusion + Calibration + Thresholding to generate final predictions.

Project Duration

March 12, 2025 - May 28, 2025

---

🚀 Pipeline Overview

This project consists of the following stages:

1. Image Preprocessing

• Orientation normalization tailored for salamander images
• 5-Crop Test-Time Augmentation (TTA)

2. Feature Extraction

• MegaDescriptor: Extracts global embeddings
• ALIKED: Local keypoint-based similarity
• EVA02: Pretrained global descriptor based on CLIP

3. Feature Fusion

• Combines MegaD and ALIKED outputs using the WildFusion calibration module
• WildFusion normalizes scores via calibration

4. Final Similarity Computation

• Combines WildFusion similarity and EVA02 cosine similarity using weighted sum
• Can be extended with a Fusion MLP in the future

5. Individual Identification

• Threshold-based binary classification:
  • Above threshold → existing individual
  • Below threshold → classified as new_individual

---

💡 Key Features

component	Description
MegaDescriptor	Global feature extractor using Timm
ALIKED	Local keypoint matcher
WildFusion	Fusion and calibration of two matchers
EVA02	Large-scale pretrained vision model (CLIP-based)
Rerank Cascade	Improves accuracy via top-k candidate reranking
Species-specific Strategy	Custom handling for special cases (e.g., salamanders)

---

📁 Project Structure

AnimalCLEF2025/
│
├── main.py                  # Main entry point
├── config.py                # Configuration constants
├── sample_submission.csv    # Final predictions for submission
│
├── src/
│   ├── dataset.py           # Dataset loading and sampling
│   ├── matcher.py           # Mega, ALIKED, EVA02 matcher definitions
│   ├── fusion.py            # WildFusion module
│   ├── fusion_head.py       # Fusion MLP head
│   ├── transforms.py        # Preprocessing and TTA strategies
│   └── utils.py             # Utility functions
│
├── assets/
│   └── pipeline.png         # Pipeline diagram

---

🔧 How to Run

1. Install dependencies:

pip install -r requirements.txt

2. Set up Kaggle API (download kaggle.json from your Kaggle account settings and place it in ~/.kaggle/):

mkdir -p ~/.kaggle
mv ~/Downloads/kaggle.json ~/.kaggle/
chmod 600 ~/.kaggle/kaggle.json

3. Run the main script:

python main.py