Abstract

Dense mapping has been a very active field of research in recent years, promising various new application in computer vision, computer graphics, robotics, etc. Most of the work done on dense mapping use low-level features, such as occupancy grid, with some very recent work using high-level features, such as objects. In our work we use an active patch model to learn the prominent, primitive shapes commonly found in indoor environments. This model is then fitted to coming data to reconstruct the 3D scene. We use Gauss-Newton method to jointly optimize for appearance reconstruction error and geometric transformation differences. Finally we compare our results with Kinect Fusion [6]. © Springer International Publishing Switzerland 2015.