Resumen
Occlusions are a set of pixels, which are visible in a single frame of two sequential frames in a video. Finding occlusions is of great importance in the field of computer vision. Precise detection of occlusions will improve the accuracy of many video processing methods, such as: frame interpolation, optical flow calculation, color propagation etc. The majority of existing methods are based on optimization of an energy function, which is computationally expensive. It is also worth noting that accurate estimation of occlusions is hard with no information about movement between frames, and knowledge of occlusions during optical flow estimation allows the algorithm to avoid wrong correspondences between pixels of frames. Taking this into consideration we present a novel method of occlusion detection based on PWC-net, an optical flow calculation algorithm. The key idea is to construct a pyramid of features with different resolutions for frame processing. This way of processing originates from a common computer graphics technique and is widely adopted. We also performed a comparison of our method with 15 similar methods on the MPI-Sintel dataset.