Radiance fields-based 3D surface reconstruction

Summary

The accurate 3D reconstruction of complex datasets remains an open research challenge, particularly in the case of scenes characterized by vegetation, transparent materials, or non-Lambertian surfaces, which often lead to poor performance of standard multi-view stereo (MVS) techniques. This issue is especially relevant in the cultural heritage field, where many objects and environments such as monumental trees, historic gardens, glazed surfaces, and polished stone materials exhibit exactly these challenging characteristics. In this research, we investigate the use of novel radiance fields (NeRF, Gaussian splats) to generate tangible 3D mesh models especially in the case of challenging surfaces where traditional MVS techniques might fail. In constrast to pure radiance fields where the main objective is novel view synthesis, in this research we focus more on the use of such methods to directly create surface meshes, which can then be used for other purposes such as metric archive or HBIM.

Lead

Dario Billi

Collaborators

• Chaimaa Delasse (ICube TRIO & IAV Hassan II)
• I Gede Mahendra Dharmawiguna (Monash University)