ROSEFusion

Fig. 1. We introduce ROSEFusion, a depth-only online dense reconstruction which is stable and robust to highly fast camera motion. Built upon the volumetric depth fusion framework, our method solves the highly nonlinear optimization problem of fast-motion camera tracking using random optimization. In this example, the depth camera moves at a speed of 2m/s in average and up to 3.6m/s, leading to severe motion blur in RGB images. This sequence would break most state-of-the-art online reconstruction methods. Thanks to our novel particle filter optimization with swarm intelligence, our method is able to fuse the depth maps (resilient to motion blur) accurately, attaining a satisfying reconstruction quality. See the planar walls and the correct overall layout; the imperfect local geometry was mainly due to incomplete depth scanning. The tracked camera trajectory is visualized and color-coded with camera speed.