Skip to content

Underwater video analysis · AI · Open source

Community Fish Detector

A community-built AI model trained to detect one class, 'fish', and generalise across marine environments and species.

Try it out GitHub Dataset

Video from Global Finprint Project YouTube

Through a community-driven effort, we standardised the publicly available, fragmented fish-imaging datasets into a single unified dataset, and used it to train a "universal fish model" that aims to detect fish in any underwater video or image, regardless of camera, depth, environment or species.

The problem

Fragmented AI approaches

  • • Every project trains its own model from scratch.
  • • Useful datasets/models exist, but in incompatible formats.
  • • Standardising data is costly; few groups can do it alone.

The approach

Community-led standardisation

  • • Contributors each adopt one dataset and convert it to COCO format.
  • • The cost of standardisation is split across many groups.
  • • Dataset, model and tooling are openly shared.

The Dataset

Each contributor adopted one open dataset and converted it to a common COCO-format schema, with a single fish class. The 17 sources span BRUVS, ROVs, deep-sea cameras, river-herring weirs, coral reefs, aquaculture pens, lab tanks and freshwater rivers, covering most of the imaging conditions a fish detector might encounter in the wild.

We aggregated the converted datasets, deduplicated labels, audited bounding boxes for consistency, and split the result into training, validation and held-out evaluation sets, with the held-out split used to measure how the model generalises beyond the data it was trained on.

The dataset is openly hosted on LILA BC: lila.science/datasets/community-fish-detection-dataset .

17

contributors

17

datasets processed

2M

images, 935K annotations

The Model

We trained two single-class detectors on the CFD corpus and release both as open weights:

RF-DETR · Default

Permissive license

A DETR-style transformer detector trained on the same corpus, released under a permissive open-source license (Apache 2.0). You can use, fine-tune and ship it inside commercial or closed-source projects without copyleft obligations. This is why we recommend RF-DETR as the default model.

YOLOv12x

Faster, AGPL-licensed

A real-time anchor-free detector, faster at inference and lighter to deploy on laptops and edge devices. Inherits Ultralytics' AGPL-3.0 license, a strong copyleft that requires any work using it (including networked services) to be open-sourced under the same terms. Pick YOLOv12x when speed matters and you're comfortable with the AGPL obligations.

Training, evaluation and inference code is openly developed on github.com/filippovarini/community-fish-detector , including dataset preparation scripts, training configs and the held-out evaluation protocol.

79%

Average AI Precision (AP@50) on held-out val data

1

class: "fish", across every domain

Example detections across some of the imaging modalities CFD was trained on.

CFD detections, Coral reef
Coral reef
CFD detections, Deep sea
Deep sea
CFD detections, Pelagic
Pelagic
CFD detections, BRUVS
BRUVS
CFD detections, Freshwater
Freshwater
CFD detections, Turbid water
Turbid water

Evaluation against SOTA models

Comparison of CFD (YOLOv12x) with MegaFishDetector (MFD, Yang et al. 2023) on held-out evaluation datasets. Results in AP@50.

Dataset CFD MFD
fathomnet* 54 46
deepfish* 85 63
noaa_puget* 26 1
fishtrack* 52 16
brackish 66 1
coralscapes 32 1
deep_vision 88 23
f4k 67 1
kakadu 81 2
marine_detect 70 8
river_herring 83 2
project_natick 37 0
roboflow_fish 69 11
salmon_cv 94 11
torsi 79 12
zebrafish 99 65

* Dataset MFD was trained on.

Get started

Free to use, openly developed

The model, dataset and tooling are all open. Try the detector in your browser, pull the weights from GitHub, or download the dataset from LILA.

🚀

Try it out

Run CFD on your images in the Hugging Face Space.

💻

GitHub

Model weights, training code and documentation.

📊

Dataset

The CFD Dataset on LILA: 2M images, 935K annotations.

Questions? Contact us.

Citation

Cite our work

Varini, F. et al. (2026). Community Fish Detector. GitHub. github.com/filippovarini/community-fish-detector 
@misc{varini2026cfd,
  title  = {Community Fish Detector: a community-built single-class fish detector
            generalising across marine imaging domains},
  author = {Varini, F. and contributors},
  year   = {2026},
  howpublished = {\url{https://github.com/filippovarini/community-fish-detector}}
}

Model and dataset are openly available. Submit issues on GitHub, we usually respond within a week.

Contributors

Built by many hands

The Community Fish Detector exists thanks to the efforts of Filippo Varini, Dan Morris, Sonny Burniston, Océane Boulais, Kevin Barnard, Laura Chrobak, Alexander Merdian-Tarko, Devi Ayyagari, Mona Dhiflaoui, Joshua Chen, Gerard Calvo Bartra, Kalindi Fonda, Levi Veevee Cai, Giorgio De Pertis, Chris Jackett, Aditya Shirvalkar and Adrian Ibanez.

Get involved

We're building v2

We're looking for contributors to expand and improve the dataset, retrain the model and grow the working group. If you maintain a marine-imaging dataset, want to help curate one, or want to collaborate on the next iteration of the universal fish model, reach out.

Reach out to collaborate