Deep Learning for Particle Tracking

Master’s thesis: Deep Learning for Particle Tracking

Although everyone has come into contact with a raindrop, few of us are aware of the plethora of physical processes and complex interactions that are involved in creating one. Many of these processes are not yet fully understood due to the large range of scales involved [1,2]. To shed light on this question, we have built a sophisticated particle tracking setup using three high-speed cameras at the environmental research station Schneefernerhaus [3,4], situated near Germany’s highest mountain Zugspitze. Detecting the particles from the recorded images is a non-trivial image recognition problem that constitutes a major computational bottleneck in our current process. In this project we want to employ modern deep-learning based object detection systems to speed up the particle detection process and potentially improve its accuracy when image quality is suboptimal. The scope of the project includes creating a suitable dataset for training and evaluation, adapting and optimizing a number of state-of-the-art object detection systems to the custom image types (including, e.g. self-supervised pre-training) and comparing their performance in terms of speed and accuracy to the existing solution.

Particle Tracking Setup

Literature

[1] A. Pumir, et al., Annu. Rev. Condens. Matter Phys., 7, 141-70 (2016)
[2] G. P. Bewley, E.-W. Saw, E. Bodenschatz, New J Phys., 15, 083051 (2013)
[3] S. Risius, H. Xu, F. DiLorenzo, et al., Atmos. Meas. Tech., 8, 3209 (2015)
[4] H. Siebert, R. A. Shaw, et al., Atmos. Meas. Tech., 8, 3219-3228 (2015)

Requirements

  • Experience with deep learning using Python and PyTorch
  • Completed lecture “Deep Learning” (B.Inf.1237)
  • Good mathematical understanding (in particular statistics and linear algebra)

Contact

Alexander Ecker and Jan Molacek

Neural Data Science Group
Institute of Computer Science
University of Goettingen