Video Semantic Segmentation

Semantic segmentation is a task of dividing an image into semantically meaning components. It is one of a key in the field of a computer vision and has many practical applications. It can be solved by the pixel-wise picture classification to one of the predefined classes. In the work presented here, the existing ways of solving problems of image and video semantic segmentation are explored. Most of them ignore the fact of the dependence of frames in the video sequence among themselves. Those who take this property into account use it to accelerate calculations by reducing the number of frames to be processed by the segmentation model and using the results received for them to construct predictions for the remaining ones frames also. This work is partially contrasted with this approach. The main tasks are to study segmentation models and improve the quality of their work, adding an interpolator to the process of building a prediction that aggregates information from neighboring frames. The article shows that this interpolation can be one of the ways to smooth and refine the edges of objects. Also, in the process of implementing this work, a data set containing the next neighboring frames for each in the CamVid collection was obtained.