Urchinbot Dataset - Sea Urchin Object Detection and Classification

An object detection dataset used for training, validating and evaluating sea urchin detection models. This dataset contains 9,872 images and over 44,000 annotations belonging to three urchin species from a variety of locations around New Zealand and Australia.

Complete_urchin_dataset.csv contain a full list of images in the dataset and the corresponding bounding boxes and additional metadata, including image source, campaign/deployment names, latitude, longitude, depth, altitude, time stamps and more. High_conf_clipped_dataset.csv is a preprocessed version of the complete dataset that has removed annotations with low annotators' confidence scores (< 0.7), removed annotations that are flagged for review and clipped all bounding boxes to fit within the bounds of the images.

Run the download_images.py script to download all the images from the URLs in the csv files.

Labels.zip (YOLOv5 formatted txt bounding box label files), yolov5_dataset.yaml (YOLOv5 dataset configuration file) and train/val/test.txt (training, validation and test splits) can be used to train a YOLOv5 object detection model on this dataset. 

See https://github.com/kraw084/Urchin-Detector for code, models and more documentation relating to this dataset.