Enhanced 3D Capture for Room-sized Dynamic Scenes with Commodity Depth Cameras Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 19, 2019
  • Dou, Mingsong
    • Affiliation: College of Arts and Sciences, Department of Computer Science
  • 3D reconstruction of dynamic scenes can find many applications in areas such as virtual/augmented reality, 3D telepresence and 3D animation, while it is challenging to achieve a complete and high quality reconstruction due to the sensor noise and occlusions in the scene. This dissertation demonstrates our efforts toward building a 3D capture system for room-sized dynamic environments. A key observation is that reconstruction insufficiency (e.g., incompleteness and noise) can be mitigated by accumulating data from multiple frames. In dynamic environments, dropouts in 3D reconstruction generally do not consistently appear in the same locations. Thus, accumulation of the captured 3D data over time can fill in the missing fragments. Reconstruction noise is reduced as well. The first piece of the system builds 3D models for room-scale static scenes with one hand-held depth sensor, where we use plane features, in addition to image salient points, for robust pairwise matching and bundle adjustment over the whole data sequence. In the second piece of the system, we designed a robust non-rigid matching algorithm that considers both dense point alignment and color similarity, so that the data sequence for a continuously deforming object captured by multiple depth sensors can be aligned together and fused into a high quality 3D model. We further extend this work for deformable object scanning with a single depth sensor. To deal with the drift problem, we designed a dense nonrigid bundle adjustment algorithm to simultaneously optimize for the final mesh and the deformation parameters of every frame. Finally, we integrate static scanning and nonrigid matching into a reconstruction system for room-sized dynamic environments, where we prescan the static parts of the scene and perform data accumulation for dynamic parts. Both rigid and nonrigid motions of objects are tracked in a unified framework, and close contacts between objects are also handled. The dissertation demonstrates significant improvements for dense reconstruction over state-of-the-art. Our plane-based scanning system for indoor environments delivers reliable reconstruction for challenging situations, such as lack of both visual and geometrical salient features. Our nonrigid alignment algorithm enables data fusion for deforming objects and thus achieves dramatically enhanced reconstruction. Our novel bundle adjustment algorithm handles dense input partial scans with nonrigid motion and outputs dense reconstruction with comparably high quality as the static scanning algorithm (e.g., KinectFusion). Finally, we demonstrate enhanced reconstruction results for room-sized dynamic environments by integrating the above techniques, which significantly advances state-of-the-art.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Fuchs, Henry
  • Izadi, Shahram
  • Dunn, Enrique
  • Cham, Tat-Jen
  • Frahm, Jan-Michael
  • Lastra, Anselmo
  • Doctor of Philosophy
Degree granting institution
  • University of North Carolina at Chapel Hill Graduate School
Graduation year
  • 2015
Place of publication
  • Chapel Hill, NC
  • There are no restrictions to this item.

This work has no parents.