Enhancing defocused particle tracking with modified detection and localization
M.J. Yamakaitis(1), C. Fort(1), Y. Sawalhi(1), J.R. Rosen(2), P.M. Bardet(1)
1: Dept. of Mechanical and Aerospace Engineering, The George Washington University, Washington, DC, United States
2: School of Public Policy and Urban Affairs, Northeastern University, Boston, MA, United States
DOI:
A recently published convolutional neural network architecture designed for conventional multicamera image-based particle tracking is adapted to a three-dimensional single-camera imaging domain. Originally demonstrated for high particle densities where particle images are well approximated by Gaussian profiles only a few pixels wide, it is extended to `defocus' particle tracking, where particle images are much larger and have arbitrarily complex shapes that encode depth. The single network performs simultaneous detection and localization in all three spatial dimensions for a single-image snapshot. We demonstrate simultaneous precision and recall of over 95% on previously published benchmark datasets, exceeding the performance of previous techniques. We extend the published dataset and test detection on images with 45 times more particles per pixel.
