BOR$^2$G: Building Optimal Regularised Reconstructions with GPUs (in Cubes)

Abstract

This paper is about dense regularised mapping using a single camera as it moves through large work spaces. Our technique is, as many are, a depth-map fusion approach. However, our desire to work both at large scales and outdoors precludes the use of RGB-D cameras. Instead, we need to work with the notoriously noisy depth maps produced from small sets of sequential camera images with known inter-frame poses. This, in turn, requires the application of a regulariser over the 3D surface induced by the fusion of multiple (of order 100) depth maps. We accomplish this by building and managing a cube of voxels. The combination of issues arising from noisy depth maps and moving through our workspace/voxel cube, so it envelops us, rather than orbiting around it as is common in desktop reconstructions, forces the algorithmic contribution of our work. Namely, we propose a method to execute the optimisation and regularisation in a 3D volume which has been only partially observed and thereby avoiding inappropriate interpolation and extrapolation. We demonstrate our technique indoors and outdoors and offer empirical analysis of the precision of the reconstructions.