In this thesis, we have developed a system for creating synthetic sequences suitable for use in visual tracking to evaluate or learn tracking algorithms. The system is designed as an extension for the 3D modelling software Blender and allows for parametrisation of dynamics of different scene features, e.g. camera dynamics, lighting. This way of generating sequences also allows for quick generation of accurate labels needed for evaluation or learning.
To demonstrate the applicability of our approach, we created five scenes using which we generated 30 sequences and evaluated four reference trackers on them. By changing the parameters of the sequences, we can demonstrate certain limitations of each tracker that can be linked to their design.
|
