Creating indoor space model based on single RGB image with the segmentation of surface free to move

This paper suggests using existing neural networks and photogrammetry methods to build a point cloud model of a room and highlight floor surfaces. The work uses neural networks for segmentation: Boundary-guided Context Aggregation Network, Deep Convolutional Neural Networks with a fully connected Conditional Random Field, Spatial information Guided Convolutional Network and Swin Transformer with Shifted Windows. Also work uses DPT-L for depth prediction. The proposed ensemble demonstrated the viability of the concept. Three shortcomings were also identified: high peak load on hardware, sensitivity to the color palette of the input image, and the need for fine-tuning. When solving or reducing the impact of identified problems, the ensemble can be used for remotely controlled land-based unmanned vehicle.