Bachelorarbeit BCLR-2019-76

Yang, Haining: Flow prediction meets flow learning: combining different learning strategies for computing the optical flow.
Universität Stuttgart, Fakultät Informatik, Elektrotechnik und Informationstechnik, Bachelorarbeit Nr. 76 (2019).
61 Seiten, englisch.

Optical flow estimation is an important topic in computer vision. The goal is to computethe inter-frame displacement field between two consecutive frames of an image sequence. In practice, optical flow estimation plays a significant role in multiple application domains including autonomous driving and medical imaging. Different categories of methods exist for solving the optical flow problem. The most common technique is based on a variational framework, where an energy functional is designed and minimized in order to calculate the optical flow. Recently, other approaches like pipeline-based approach and learning-based approach also attract much attention. Despite the great advances achieved by these algorithms, it is still difficult to find an algorithm that can perform well under all the challenges, e.g. lightning changes, large displacements, and occlusions. Hence, it is worth combining different algorithms to create a new approach that can combine their advantages. Inspired by this idea, in this thesis we select two top-performing algorithms PWC-Net and ProFlow as candidate approaches and conduct a combination of these two algorithms. While PWC-Net performs generally well in the estimation of non-occluded areas, ProFlow can especially provide an accurate estimation for the occluded areas. Thereby, we expect that the combination of these two algorithms can yield an algorithm that performs well in both occluded and non-occluded areas. Since ProFlow is a pipeline approach, we first integrate the PWC-Net in the ProFlow pipeline, then evaluate the new created pipeline PWC-ProFlow based on the MPI Sintel and KITTI 2015 benchmarks. Contrary to our expectations, the newly created algorithm does not exceed the candidate methods PWC-Net and ProFlow on either benchmark. Through the analysis of the evaluation results, we explore the problems hidden in the PWC-ProFlow pipeline that can lead to its underperformance, and organize some modification ideas. Based on these ideas, we propose six new pipelines with the purpose of improving the estimation accuracy of PWC-ProFlow. All the new generated pipelines are also evaluated on the Sintel and KITTI benchmarks. The experiment results demonstrate that all the modifications created achieve great improvements on both datasets compared to PWC-ProFlow. Further, all of them also outperform the ProFlow pipeline on both benchmarks. Compared to PWC-Net, one modification exceeds PWC-Net on the KITTI dataset, however, all our modifications achieve a better performance on the Sintel dataset, in particular, one modification presents a significant improvement with a more than 10% lower average endpoint error on the Sintel dataset.

Volltext und
andere Links
Abteilung(en)Universität Stuttgart, Institut für Visualisierung und Interaktive Systeme, Visualisierung und Interaktive Systeme
BetreuerBruhn, Prof. André S; Richter, Maurer, Daniel
Eingabedatum19. Februar 2020
   Publ. Informatik