ИСТИНА

The use of convolutional neural networks in image processing tasks often allows achieving significantly better results in comparison with traditional methods. For the problem of calculating a real-time depth map, the state-of-the-art method is the MiDaS method - a convolutional neural network trained on a large dataset that is not publicly available. However, this approach does not include the ability to use information from neighboring frames which can improve the prediction. The method proposed in this paper uses the depth maps generated by MiDaS for several frames of a video sequence and their further refinement, which makes it possible to achieve an improvement in quality without a significant decrease in the algorithm performance.