The extraction of information considering the movement of objects in an image sequence becomes more and more an important problem, amongst others in the area of computer vision. In most cases, a displacement vector field between two consecutive frames of an image sequence should be computed which is often called optical flow. Several methods and approaches have been introduced to compute this optical flow which is why some optical flow benchmarks have been created to evaluate them. These benchmarks contain synthetic and non-synthetic data like the Middlebury Benchmark, synthetic data from an animated short film like the MPI-Sintel Benchmark or real-world data collected by an autonomous driving car like the KITTI Vision Benchmark Suite. However, the difficulties of these benchmarks haven't been investigated so far which is why different metrics should be developed in this thesis to evaluate them. These address image statistics, optical flow statistics, illumination changes, the type of movement and egomotion in stereo scenes. First of all, they are applied on the training data with ground truth flow to estimate afterwards the difficulty of the testing data. The benchmarks which are used are the three mentioned before.