Volumetric Fusion

Final project for the 3D Scanning & Spatial Learning practical course from the Chair of Computer Graphics and Visualization at Technische Universität München in the WS 2019/20.
The course is held by Dr. Justus Thies and our group is supervised by Aljaž Božič.

Overview

A multi-view RGB-D Capture Setup is built using 4 Intel® RealSense™ Depth Camera D415 to perform real-time reconstructions of a moving scene.

Steps

• Image preprocessing
  • Depth filtering
  • Background subtraction
  • Visual hull computation
• Non-rigid shape reconstruction
  • Sparse voxel grid
  • Deformation graph construction
• Non-rigid tracking
  • Projective depth ICP
  • Global sparse correspondences

Data

• 4 Intel® RealSense™ Depth Camera D415

Literature

• Motion2Fusion: Real-time Volumetric Performance Capture [Dou et al.]
• Fusion4D: Real-time Performance Capture of Challenging Scenes [Dou et al.]
• DynamicFusion: Reconstruction and Tracking of Non-rigid Scenes in Real-Time [Newcombe et al.]
• RealSense Multiple Camera WhitePaper
• Volumetric Capture
• Voxel grid TSDF
• OpenGL Compute Shader

Github Repos

• tsdf Fusion
• tsdf Fusion in Python
• Elastic Fusion
• ReFusion
• KinectFusion
• ReFusion
• Kintinuous - Real-time large scale dense visual SLAM system