A deep neural network for high‐throughput measurement of functional traits on museum skeletal specimens

Increasingly, natural history museum collections are being used to generate large-scale morphological datasets address a range of macroecological and macroevolutionary questions. One challenge this approach is that large numbers individuals either from single species or taxonomically broad sets may be necessary characterize morphology at the relevant spatial, phylogenetic temporal scales. We present ‘Skelevision’, method for rapidly handling, photographing measuring skeletal specimens with computer vision uses deep neural network segment photographs into individual bones, identify measure functional aspects those bones. demonstrate scale what feasible Skelevision by estimating 11 traits different bones 12,450 bird spanning 1,882 passerines (~32% all passerine diversity). quantify accuracy estimates comparing them handmade measurements 174 115 across 79 genera span 59 families. precise, mean standard deviation 0.86 mm repeated independent extremely accurate, RMSE 0.89 when compared measurements. There minimal signal in measurement error (mean Pagel's λ = 0.13), robust variation degree which remain articulated. This has several important advantages over traditional methods building (e.g. long-term field-based operations specimens). First, new only requires collection photographs, can then measured automatically, effectively instantaneously, network. significant departure time skill required hand. Second, repeatable. Third, even as dataset photographed expands, amount annotation data needed on using will fixed done without re-capturing images.