Methodology:

Data was collected as part of expedition PS118, cruises 69-1 and 6-9, of the RV Polarstern in 2019 [1] using the OFOBS [2] and manually labelled for use in benthic community analysis [3]. These labels were then condensed into 25 morphotypes, with annotations converted to COCO-formatted bounding boxes, for use in object detection model development. Data was split into training, validation, and test sets based on substrate, depth, seafloor inclination.

Imagery in this dataset is a subset of imagery collected during the expedition PS118 [1], available on PANGEA [2]. Dataset annotations are a subset of those present in [4] for use in benthic community analysis [3]. This dataset was used for the development of an object detection model capable of automated benthic organism detection [in-prep]. For model weights, see Related URLs.

Some original source images [4] were not comprehensively annotated, e.g. due to distortion. For use in object detection model training, the unlabelled regions were cropped, resulting in images of varying dimensions (average size = 3,364×4,545px).

Data collection:

Data was collected using the OFOBS, a top-down towed camera system. Original labelling for benthic community analysis was performed in Inkscape v1.1. Resulting SVG files were then converted using a custom script file to JPGs with corresponding COCO bounding box JSON using Python v3.12.8. Converted bounding boxes were then manually edited (resized etc) using LabelMe v5.8.1, then converted back to COCO format.

Data quality:

All bounding boxes were manually checked after conversion from SVG format. Class labels have been reviewed by ecologists at BAS for accuracy. Given the high densities of organisms in the dataset, the prevalence of small-bodied taxa, and the well documented issues of fatigue and subjectivity in manual annotation processes for benthic imagery, it is likely some valid organisms were omitted from the ground truth.