Collaborative Large-Scale Dense 3D Reconstruction with Online Inter-Agent Pose Optimisation.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Robotics

Background:

• Reconstructing large 3D scenes is crucial but time-consuming, increasing the risk of scene changes.
• Joining smaller captured sub-scenes is challenging due to novel poses and tracking drift.
• Recent advances in real-time globally consistent reconstruction and visual-inertial odometry have improved sub-scene capture.

Purpose of the Study:

• To present the first system enabling interactive, collaborative reconstruction of dense, voxel-based 3D models of entire buildings.
• To overcome the limitations of traditional methods requiring specialized hardware and significant capture time.
• To leverage recent advancements in relocalization and drift reduction for practical large-scale 3D scene capture.

Main Methods:

• Utilizing real-time globally consistent reconstruction systems for on-the-fly surface re-integration.
• Employing advanced visual-inertial odometry to minimize tracking drift during live reconstruction.
• Integrating online-trained regression forest-based relocalizers for robust pose estimation.

Main Results:

• Demonstrated the first system for collaborative, interactive reconstruction of dense, voxel-based building models.
• Achieved reconstruction of entire houses or labs in under 30 minutes using consumer-grade hardware.
• Significantly reduced the cost and time associated with large-scale 3D scene capture.

Conclusions:

• The developed system successfully enables collaborative, efficient, and cost-effective 3D reconstruction of large-scale environments.
• It overcomes traditional barriers by integrating recent advancements in computer vision and robotics.
• This facilitates widespread adoption of detailed 3D scene modeling for various applications.