Single camera estimation of microswimmer depth with a convolutional network.

Abstract

A number of techniques have been developed to measure the three-dimensional trajectories of protists, which require special experimental set-ups, such as a pair of orthogonal cameras. On the other hand, machine learning techniques have been used to estimate the vertical position of spherical particles from the defocus pattern, but they require the acquisition of a labelled dataset with finely spaced vertical positions. Here, we describe a simple way to make a dataset of <i>Paramecium</i> images labelled with vertical position from a single 5 min movie, based on a tilted slide set-up. We used this dataset to train a simple convolutional network to estimate the vertical position of <i>Paramecium</i> from conventional bright field images. As an application, we show that this technique has sufficient accuracy to study the surface following behaviour of <i>Paramecium</i> (thigmotaxis).