System and method for training a neural network to fill gaps between scan points in images and to de-noise point cloud images

A system for training a neural network, the system comprising:

• a processor;
• and
• a memory, the memory including instructions that, upon execution by the processor, cause the processor to be operable to:
• perform, using a laser scanner having a color camera, a scan of an environment, the laser scanner being configured to measure a plurality of three-dimensional (3D) coordinates based at least in part on the speed of light in air;
• generate scan point cloud data from the scan, the scan point cloud data comprising scan points and gaps between the scan points;
• generate a training data set of point clouds, the training data set comprising pairs of closed surfaces point clouds and non-closed surfaces point clouds;
• for each of the closed surfaces point clouds and the non-closed surface point clouds,
• generate a two-dimensional (2D) image by rendering a 3D scene, wherein the 2D image for the non-closed surface point clouds include gaps, and wherein the 2D image for the closed surface point clouds are free of gaps, wherein the 2D image comprises color information associated with the plurality of target objects and the plurality of background objects, the color information being obtained from the color camera;
• train the neural network, to generate a trained neural network, using the 2D images for the non-closed surface point clouds and the 2D images for the closed surface point clouds, the trained neural network being trained to identify one or more gaps in one or more surfaces of the plurality of target objects rather than in the plurality of background objects and fill in the one or more gaps in the one or more surfaces of the plurality of target objects of the 2D image for the non-closed surface point clouds;
• fill, using the trained neural network, the one or more gaps in the one or more surfaces of the plurality of target objects between the scan points of the 2D image for the non-closed surfaces point cloud based at least in part on the color information;
• and
• de-noise, using the trained neural network, the scan point cloud data to generate a closed surface image of the scan point cloud data.