SC²S Colloquium - August 10, 2017

From Sccswiki
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to navigation Jump to search
Date: August 10, 2017
Room: 02.07.023
Time: 16:00 pm, s.t.

Simon Griebel: Optimizing Load Balancing of Heterogeneous Particle Simulations on Heterogeneous Systems

In this thesis improvements to the k-d-tree based load balancing algorithm of the molecular dynamics simulation framework ls1-MarDyn will be presented. Its central part, the load estimation function, was solely based on heuristics, so for this thesis it was replaced with a function that actually uses time measurements for the estimation. These measurements then made it easier to expand the algorithm, so that it would correctly work for more than one particle type and on heterogeneous systems. Additionally there was a look at how certain modifications to the splitting rules of the k-d-tree-nodes can influence performance.

San Yu Huang: Semi-Global Matching in Stereo and Epipolar Flow Estimation

In the field of computer vision, stereo and flow estimation are classic topics in real-time appli- cations. The purpose of this thesis is to implement stereo and epipolar flow estimation with Semi-Global Method. At the current stage, a few publications apply this method to stereo esti- mation, but rarely in flow estimation. The pixel-wise cost function with Census transformation is involved in both scenarios. During cost computation, potential targets are selected based on mapping of corresponding epipolar geometry. In particular, the geometrical representation gives 1-D pixel-wise optimization in flow estimation. Afterward, aggregating cost by sweeping the image domain is performed. Equal-weighted in a number of paths results in better accu- racy and lower runtime. The implementation and corresponding mathematical model are also presented step by step. The disparity map of stereo matching is evaluated by KITTI database and journal papers. The baseline is achieved and to be extended to flow estimation. Finally, accelerating the program on the GPU by using CUDA library is carried out.