Real-Time Approach to Neural Network-Based Disparity Map Generation from Stereo Images

Generating depth maps using mono- or stereo- images is a topic of active research. This paper is dedicated to study of different methods of depth and disparity maps generating. It includes analysis of existing methods of depth maps generating and investigation and improvement of the real-time neural-network based method AnyNet. Our approach AnyNet-M related to the models which are make prediction stage-by-stage and increase quality of disparity estimation over time. Firstly, proposed method was tested on KITTI Stereo Dataset and custom OpenTaganrog dataset with images of 1242 × 375 resolution. The method was proved to be real-time with approximately 38 FPS using one TeslaV100 GPU. Secondly, quality of proposed model was tested and reached 5% of Average 3-Pixel Error on KITTI dataset. Finally, it was integrated to Robotic operation system for further use as part of the navigation systems of unmanned vehicles.