Repeatable Semantic Reef-Mapping through Photogrammetry and Label-Augmentation

In an endeavor to study natural systems at multiple spatial and taxonomic resolutions, there is urgent need for automated, high-throughput frameworks that can handle plethora of information. The coalescence remote-sensing, computer-vision, deep-learning elicits a new era in ecological research. However, complex systems, such as marine-benthic habitats, key processes still remain enigmatic due the lack cross-scale automated approaches (mms kms) community structure analysis. We address this gap by working towards scalable comprehensive photogrammetric surveys, tackling profound challenges full semantic segmentation 3D grid definition. Full (where every pixel classified) extremely labour-intensive difficult achieve using manual labeling. propose label-augmentation, i.e., propagation sparse labels, accelerate task photomosaics. Photomosaics are synthetic images generated from projected point-of-view model. navigation sensors (e.g., diver-held camera), it repeatably determine slope-angle map. show especially important topographical settings, prevalent coral-reefs. Specifically, we evaluate our approach on benthic three different environments challenging underwater domain. Our label-augmentation shows human-level accuracy photomosaics labeling 0.1%, evaluated several measures. Moreover, found definition leveler improves consistency community-metrics obtained occlusions topology (angle distance between objects), were able standardise transformation with two percent error size measurements. By significantly easing annotation process standardizing present mapping methodology enabling change-detection, which practical, swift, cost-effective. workflow enables repeatable surveys without permanent markers specialized gear, useful research monitoring, code available online. Additionally, release Benthos data-set, fully manually labeled oceanic over 4500 segmented objects computer-vision marine ecology.