Globally Consistent Dense Real-Time 3D Reconstruction from RGBD Data

Kurzfassung

In this work, we present a dense 3D reconstruction framework for RGBD data that can handle loop closure and pose updates online. Handling updates online is essential to get a globally consistent 3D reconstruction in real-time. We also introduce fused depth maps for each keyframe that contain the fused depths of all associated frames to greatly increase the speed for model updates. Furthermore, we show how we can use integration and de-integration in a volumetric fusion system to adjust our model to online updated camera poses. We build our system on top of the InfiniTAM framework to generate a model from the semi-dense, keyframe based ORB SLAM2. We extensively evaluate our system on real world and synthetic generated RGBD data regarding tracking accuracy and surface reconstruction.