#DeOlhoNosCorais: a polygonal annotated dataset to optimize coral monitoring

The first version of the #DeOlhoNosCorais dataset was built in collaboration with the Laboratory of Marine Ecology (LECOM) at the Federal University of Rio Grande do Norte (UFRN). The dataset contains images extracted from social networks with the hashtag “#DeOlhoNosCorais” and images provided by the LECOM. It contains 1411 images and 21 classes. Each image has two labels, its major class and the segmentation map. The labeling was performed by the LECOM researchers using the Labelme tool. The duplicate images were removed using the FiftyOne tool. 

One of the objectives of the dataset is to measure the viability to use machine learning to speed up the processing of analyses of the images extracted from the Instagram. So, the extra images taken by divers from the researcher group were used to augmented the train set. The Instagram images were divided into train, validation, and test accordingly to their date. The rules for the division was:

• Train set: images dated until 12/31/2019 + extra images from LECOM-UFRN
• Validation set: images dated between 01/01/2019 to 06/30/2019
• Test set: images dated between 07/01/2019 to 08/24/2021

The test set includes an interval of 2 years, due to the drop in the number of posts during the Covid-19 pandemic.The full dataset folder contains the raw labeling using the Labelme. The experiments folder contains treated datasets for the following tasks:

• Full images classification
• Sub images classification
• Binary semantic segmentation
• Binary object detection

In addition, the experiments folder has an extra data from the Pacific Labeled Corals (PLC). The extra data contains 12 classes of the 20 present in the PLC, and the images patches were extracted centered in the labeled pixels with size of 224 x 224.