The recent commercialization of consumer-grade range cameras promises to enable almost anyone to make detailed 3d scans of their environments. However, there are two difficulties in using range data collected by these cameras to acquire detailed scene models. The first is the fidelity of the data. Consumer-grade range sensors produce range images with many errors, including high-frequency noise, quantization, and substantial low-frequency distortions. The second difficulty is the complexity of the camera trajectory that is necessary for a detailed reconstruction. In a complex scene, the operator must move the sensor along a trajectory that weaves around objects to image them from many points of view, in order to overcome occlusion. Estimating these camera trajectories given inaccurate input data is a highly challenging problem.
We took inspiration from previous research in computer graphics, computer vision, and robotics. We have developed a number of ideas and built the state-of-the-art offline scene reconstruction system.