The geometric reconstruction of real objects using short-range 3D measurements typically carried out with portable RGB-D scanners, presents several critical aspects that have not yet been fully resolved. One of the most challenging aspects is the recovery from tracking loss with respect to the forming model, especially in a real-time acquisition and reconstruction context. This paper presents a fully 3D pipeline for high-quality 3D object reconstruction. The pipeline integrates a traditional cumulative approach with robust tracking recovery solutions based on geometric deep learning and dynamic pose optimization. The solution is completely geometric, based on learned 3D feature extraction and matching. It significantly outperforms state-of-the-art reconstruction based on 2D features extracted from image data.
D2R2: Deep-Learning Based Real-Time Dense Object Reconstruction
Savardi M.;Signoroni A.
2026-01-01
Abstract
The geometric reconstruction of real objects using short-range 3D measurements typically carried out with portable RGB-D scanners, presents several critical aspects that have not yet been fully resolved. One of the most challenging aspects is the recovery from tracking loss with respect to the forming model, especially in a real-time acquisition and reconstruction context. This paper presents a fully 3D pipeline for high-quality 3D object reconstruction. The pipeline integrates a traditional cumulative approach with robust tracking recovery solutions based on geometric deep learning and dynamic pose optimization. The solution is completely geometric, based on learned 3D feature extraction and matching. It significantly outperforms state-of-the-art reconstruction based on 2D features extracted from image data.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
