A Time-Efficient Optimisation Framework for Parameters of Optical Flow Methods

Due to the increase of optical flow benchmark data, concerning both amount and resolution, learning parameters from training sequences with ground truth has become significantly more challenging in recent years. Moreover, optical flow methods are much more complex than a few years ago resulting in a larger amount of model parameters and a noticeably increased runtime. As a consequence, even optimising a small set of suitable parameters may take hours or even days which makes hand tuning infeasible. Hence, time-efficient strategies for automatic parameter optimisation become more and more important. In this context, our work addresses three important aspects. First, we provide an overview of different optimisation strategies and juxtapose them in the context of different optical flow methods and different evaluation benchmarks. Second, we focus on choosing a suitable subset of the training data to speed up the computation while still obtaining meaningful results. Finally, we also consider different strategies for distributing the evaluation on hardware infrastructures which allows to further reduce the run time. Experiments show that the proposed methodology allows to obtain good results while keeping the overall effort reasonably low.