BoostCompTrack: A multi-purpose tracking framework for salmon welfare monitoring in challenging environments

Last updated: July 29, 2025

Authors: Espen Uri Høgstedt, Christian Schellewald, Annette Stahl, Rudolf Mester

Repository: BoostCompTrack

🐟 Overview

This repository contains the codebase for BoostCompTrack, a flexible computer vision-based tracking framework designed for automated salmon welfare monitoring in industrial aquaculture net pens. Built upon BoostTrack: Boosting the similarity measure and detection confidence for improved multiple object tracking to address the unique challenges of underwater tracking. Our framework demonstrates strong performance in crowded scenes, during salmon turning, and enables monitoring of tail beat wavelength — an important welfare indicator.

📂 Project Structure

BoostCompTrack/
├── annotation_helpers/            # Tools for visualizing annotations
├── associator/                    # Object association logic
├── detector/                      # Detection modules
├── evaluation/                    # Benchmark visualization and TurnSalmon dataset formatting
├── helpers/                       # Utility functions
├── paper_helpers/                 # Scripts for generating paper figures
├── welfare_helpers/               # Modules for welfare indicators (tail beat wavelength)
├── benchmark_salmon_trackers.ipynb   # Tracker benchmarking notebook
├── extract_tailbeat_period.ipynb     # Tail beat wavelength analysis
├── track_salmon.ipynb                # Track salmon in a video

🧠 Key Features

• Pose-based tracking: Extracts bounding boxes around salmon and their body parts.
• Body-part-aware modules: Specialized modules handle occlusion and turning salmon.
• Benchmarking: Outperforms BoostTrack on two novel salmon tracking datasets.
• Tail beat analysis: Demonstrates suitability for tail beat-based welfare monitoring.

📊 Datasets

We introduce three novel datasets:

1. CrowdedSalmon – Tests robustness in dense environments.
2. TurningSalmon – Evaluates tracking during salmon turning.
3. TailbeatWavelength – For evaluating tail beat wavelength.

Datasets can be found here

📄 Citation

If you use this work in your research, please cite:

@InProceedings{Hogstedt_2025_ICCV,
    author    = {H{\o}gstedt, Espen Uri and Schellewald, Christian and Stahl, Annette and Mester, Rudolf},
    title     = {A Multi-purpose Tracking Framework for Salmon Welfare Monitoring in Challenging Environments},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2025},
    pages     = {2153--2162}
}

📬 Contact

For questions or collaborations, feel free to reach out via email.

---

BoostCompTrack is developed as part of the cAIge project, funded by the Research Council of Norway, to support the aquaculture industry in achieving continuous, automated, and precise salmon welfare monitoring.