Figure 4
From: FathomNet: A global image database for enabling artificial intelligence in the ocean
Dataset statistics contextualizing FathomNet relative to COCO, ImageNet, and iNat2017. Statistics for COCO were computed from the 2014 training split, ImageNet numbers were generated from the 2015 localization data split, and iNat2017 metrics were derived from the 2017 training data. (a) Percent of images containing a number of instances, or localizations. (b) Percent of images displaying a given number of concepts. FathomNet annotations are displayed broken down at different levels of taxonomic rank. (c) The average number of instances versus the average number of concepts over the entirety of each dataset. (d) The distribution of relative instance size. Most objects in FathomNet and iNat are small compared to ImageNet and COCO.
