Reefwise
Credits
Reefwise fish detection relies on the open datasets and models below. They are credited here as their licences require.
Datasets
Labelled Mediterranean Fish Dataset
Ignacio Catalán, Amaya Alvarez-Ellacuria (IMEDEA, CSIC-UIB), 2023
Source: doi.org/10.5281/zenodo.7534425
Licence: CC BY 4.0
Changes: the 25 original classes were merged into a single “fish” class and the “taca” class (spots on the lens) was removed. The images and annotations were used to train the fish detector.
Microsoft COCO: Common Objects in Context
Tsung-Yi Lin et al., COCO Consortium, 2014
Source: cocodataset.org
Licence: CC BY 4.0
The detection model starts from weights pre-trained on COCO. The CC BY 4.0 licence covers the annotations; COCO images remain subject to the Flickr terms of use.
Detection models
D-FINE
© The D-FINE Authors (Yansong Peng et al., USTC)
Source: github.com/Peterande/D-FINE
Licence: Apache License 2.0
Changes: fine-tuned on the Labelled Mediterranean Fish Dataset to detect fish. Uses the HGNetv2 image backbone (PaddlePaddle, Apache 2.0).
RF-DETR
© Roboflow, Inc.
Source: github.com/roboflow/rf-detr
Licence: Apache License 2.0
Changes: fine-tuned on the Labelled Mediterranean Fish Dataset to detect fish, then converted for iOS (Core ML) and Android (LiteRT).
DINOv2
© Meta Platforms, Inc. and affiliates
Source: github.com/facebookresearch/dinov2
Licence: Apache License 2.0
Image backbone used by RF-DETR.
The models above are licensed under the Apache License, Version 2.0, and distributed on an “AS IS” basis, without warranties or conditions of any kind.