A Scale-Adaptive Time-Efficient Depth Map Fusion Algorithm

Abstract

TSDF (truncated signed distance function) is a classical algorithm for real-time 3D reconstruction, which is simple and parallelizable. Still, the memory consumption and cost of time consumption limit the expansion of the reconstructed scene. To solve these problems, we proposed a new depth map fusion system. We open up a dynamically adjustable memory space, the size of which can be adaptively adjusted according to the coordinate distribution interval obtained from the GMM(Gaussian Mixture Model). When the scene exceeds the voxel coverage, a loop switch is triggered, and the mode of recycling the cache space is turned on. This system has a scene-adaptive memory management mechanism interspersed with CPU-GPU cooperative operations. It can improve the time efficiency without sacrificing the reconstruction accuracy and make it possible to reconstruct high-scale changing scenes. We have created a dataset of natural aviation scenes for extensive experiments. The experimental results show that our system is suitable for any scale of scenes. In terms of time, our method improves the time to fuse each depth map frame by order of magnitude. In evaluating the quality of the reconstructed surface, we compared our reconstructed surfaces with commercial 3D software, and the results show that more than $92 \%$ of the fractions differed from Metashape's results by less than 0.6m. In general, our system is more suitable for scenes with dynamic changes in scene scale and faster task requirements, which will promote the expansion of 3D reconstruction in practical applications.