Method and device for 3D scene reconstruction

Abstract

Embodiments of the present invention provide a method and device for 3D scene reconstruction. By including a cascaded neural network formed by cascading multiple neural networks in a hierarchical manner, firstly, multiple depth images are acquired at a lower preset resolution. The histogram corresponding to each pixel in the current processing area, combined with the current neural network to determine the sub-processing area corresponding to the next-level neural network in the current processing area and the voxel reconstruction data learned by the current neural network, and then for the next level The neural network takes the sub-processing area as the processing object, and repeats the above process on the basis of improving the resolution of the processing object; and realizes 3D scene reconstruction according to the voxel reconstruction data learned by each level of neural network. The method and device process the processing area step by step from the whole to the part in a gradual manner, which greatly simplifies the learning task of the neural network, significantly reduces the calculation cost, and has high efficiency, accuracy and practicability.

Description

technical field [0001] Embodiments of the present invention relate to the field of communication technologies, and more specifically, to a method and device for reconstructing a three-dimensional scene. Background technique [0002] High-quality reconstruction of 3D scenes has been one of the main frontiers of research in computer vision and computer graphics for many years. At the same time, the gradual popularization of consumer-grade color-depth (RGB-D) sensors (such as Microsoft Kinect and Intel RealSense) enables more novice users to perform 3D scanning of the surrounding environment, and also increases the robustness of users to 3D reconstruction algorithms. sexual needs. Despite the recent progress in 3D environment reconstruction, it is still a very challenging problem to reconstruct a high-accuracy 3D scene model from low-quality data obtained during casual scanning using consumer-grade RGB-D sensors. . [0003] Since neural networks have achieved impressive resu...

AI Technical Summary

Problems solved by technology

[0004] At present, some academic papers disclose some 3D reconstruction methods based on neural network learning, but these methods either require additional and time-consuming steps before local calculation, or need to optimize the intermediate results of 3D voxel representation globally, so that It is difficult for these methods to generate higher resolution results

Images