This repository contains an a YOLOv8 xlarge AI model trained to detect fish in underwater pictures. How to use this model: Install ultralytics: "pip3 install ultralytics" Download this model Run yolo detect from the command-line: yolo detect predict model=yolov8x_21sp_5364img.pt source= Training dataset This model is a A YOLOv8 xlarge network trained with labeled fish images acquired at OBSEA Underwater Observatory (NW Mediterranean sea). The dataset used to train the model is available here. The raw annotations without splits and with several underrepresented classes can be found here. Some pictures that have been used to train the model (around 10%) could not be shared due to licence conflicts. Technical details YOLOv8 xlarge network trained with underwater pictures. Data preprocessing Several data augmentation techniques have been used to improve the training using YOLO's built-in data augmentation options. The configuration can be found in args.yaml file. Data splitting Data has been randomly splitted in 70% training, 20% validation and 10% test. The splits are already included in the training dataset. Classes, labels and annotations The following classes have been used in training: Chromis chromis: aphia id 127000 Coris julis: aphia id 126963 Dentex dentex: aphia id 273962 Diplodus cervinus: aphia id 127051 Diplodus puntazzo: aphia id 127052 Diplodus sargus: aphia id 127053 Diplodus vulgaris: aphia id 127054 Epinephelus costae: aphia id 127034 Epinephelus marginatus: aphia id 127036 Mullus surmuletus: aphia id 126986 Muraena helena: aphia id 126303 Myliobatidae: aphia id None None Oblada melanura: aphia id 1577363 Parablennius gattorugine: aphia id 126770 Sarpa salpa: aphia id 127064 Seriola dumerili: aphia id 126816 Serranus cabrilla: aphia id 127041 Sparus aurata: aphia id 151523 Symphodus mediterraneus: aphia id 273569 Chromis chromis (back): Same as chromis chromis (splitted for training reasons, should be merged after inference) Diver: scuba diver, used mainly to prevent divers to be detected as fish Parameters The training configuration can be found at the args.yaml file Data sources Pictures where acquired by several underwater cameras deployed at OBSEA underwater observatory. Data quality Images have been manually selected to include as much variety as possible in terms of light conditions and water turbidity. Image resolution The resolution of the images in this dataset depends on the camera, it varies from 480x360 px to 2688x1520 px. Spatial coverage All pictures where taken at OBSEA underwater observatory, off-the-coast of Vilanova i la Geltrú, Spain. GPS coordinates Longitude Latitude depth 1.75257 41.18212 20 m Contact information For further technical inquiries or additional information about the annotated dataset, please contact enoc.martinez@upc.edu