Bibliography | Albrecht, Stephan: An extended analysis of difficulties and regularities in optical flow benchmarks. University of Stuttgart, Faculty of Computer Science, Electrical Engineering, and Information Technology, Master Thesis No. 4 (2017). 87 pages, english.
|
Abstract | A central problem in the field of computer vision is the extraction of movement from a sequence of images. This includes the determination of the displacement vector field between two subsequent frames. In the context of computer vision this displacement field is referred to as the optical flow. Until today many algorithms have been developed to solve the optical flow problem. In addition researchers have developed various kinds of benchmarks enabling the evaluation on performance and quality of these algorithms. The benchmarks contain real world data (KITTI [GLU12, MG15]), simple synthetic and real data (Middlebury [BSL+11]) and even rendered movies with different rendering modes (MPI Sintel [BWSB12]). However, these benchmarks do not only differ in their creation, but also in the complexity of the scenes. The reason is the different focus on different challenges of the optical flow problem. Even though these benchmarks provide a good environment for comparison, only few studies provide an overall analysis on the difficulties and regularities of these optical flow test suites. These difficulties and regularities include illumination changes, large displacements and different types of movement. One of the few works addressing such an analysis is the master thesis of Hager [Hag17]. Hager analyzed the KITTI 2015, KITTI 2012, Middlebury and MPI Sintel benchmark on the aforementioned difficulties and regularities. This thesis extends the work of Hager by presenting refined, as well as different methods and metrics to increase the interpretability of the obtained results in the different fields. Additionally, it provides a measure for researchers to help them to find image sequences containing a certain difficulty. A variational approach, based on brightness transfer functions, is introduced to measure illumination changes. The large displacement analysis is extended by a scale analysis in order to find large displacements of small objects. The movement type analysis is done using the order adaptive approach of Maurer et al. [MSB17]. The introduced metrics are tested on the training data of the benchmarks, with ground truth and computed flow, and compared to the results of Hager.
|
Full text and other links | Volltext
|
Department(s) | University of Stuttgart, Institute of Visualisation and Interactive Systems, Visualisation and Interactive Systems
|
Superviser(s) | Bruhn, Prof. Andrés; Maurer, Daniel |
Entry date | May 28, 2019 |
---|