Advances in Deep Learning-Driven Photo Identification and Meta Analysis of Cetaceans in Large Data Repositories - Supplementary Data

FinID-20 This is a dataset to demonstrate the approach detailed in Advances in Deep Learning-Driven Photo Identification and Meta Analysis of Cetaceans in Large Data Repositories. The data consist of 500 images across 20 individuals (25 images each) of the Bigg's killer whale population. The original annotations, locations, and dates have been anonymized to prevent unintended uses. The metadata contains information as to the age of each individual at the time the photograph in question was taken as well as the sex of the individual. Additionally, YOLO-style labels are provided, as well as the resulting extracted images. These extracted images are of the form that was used for training the individual classifier deteailed in the above publication. Description cropped_images This folder contains the images which have been extracted from the original source image to contain minimal background / extraneous features. These can then be used directly for the identification task. In order to provide maximum comparison capability, the train/val/test split files are also included. Images The raw images. These are the full scene images. They have one associated ID, though there may be multiple individuals in the shot. The date and location information, as well as all other meta data, have been removed from the images. notebooks This folder contains one jupyter notebook to assist in visualizing some of the YOLO labels which are present in this data repository. visualization Here are several plots regarding the distribution of the bounding box labels, such as aspect ratio distribution, center position, etc. yolo_labels This folder contains the YOLO labels for the images in the Images folder. Ethical Statement All images included in this dataset were collected by qualified professionals operating under appropriate ethical andlegal guidelines. Photographers were affiliated with licensed research organizations and possessed the necessarypermits and approvals for the collection of photographic data involving marine mammals. Image collection adhered toprotocols that ensured minimal disturbance to the animals and their natural behaviors. Additionally, data included in this release have been curated with consideration for privacy and ethical use.No personally identifying information is disclosed, and image metadata has been anonymized or obfuscated whereappropriate to prevent tracing back to sensitive locations or individuals. This dataset is provided solely forscientific and conservation-related research. Use of this dataset should remain consistent with the original intent of responsiblescientific inquiry and conservation. Usage Guidelines This dataset is intended solely for academic, conservation, and non-commercial use.Please cite the manuscript listed above when using or referencing this dataset in publications, presentations, or derivative work.Redistribution or commercial use is not permitted without written consent from the corresponding authors. For questions regarding the dataset or its usage, please contact: Alexander Barnhill: [e-mail] Bergler: [e-mail] Benchmark Results With no optimization or hyperparameter search performed, the aforementioned ID module achieves a nearest centroid classifier accuracy (NCC accuracy) of 0.72 and a precision at rank 1 (P@1) of 0.685. Dataset Nearest Centroid Classifier Accuracy (NCC Accuracy) Precision at rank 1 (P@1) FinID-20 (BASELINE) 0.720 0.685 Configuration Parameter Value Optimizer Adam (lr=1e-4, b1=0.9, b2=0.999) Model backbone EfficientNet B0 Pooling GeM (p=3) Batch Size 16 Embedding Size 512 Loss Function SubCenter ArcFace (subcenters=3) Maximum Augmentations 1 Maximum Epochs 100 Citation For use in any subsequent publications, please follow the citation details in CITATION.md.