3DGV Talk: Wenping Wang --- Studies on 3D Reconstruction.

In this talk on 3D reconstruction, I will first present a new pipeline
for scanning and reconstructing sherds (i.e. excavated ceramic fragments
on archeological sites) in large throughput, which is a long-standing
problem in digitization for archeology. Existing image acquisition
systems typically take several minutes to scan a single sherd so they
are not efficient enough for digitizing hundreds of sherds that are
typically excavated daily on an archeological site. Our image
acquisition system is capable of scanning over a thousand pieces per day
(in eight hours). Our system is not only efficient but also portal,
affordable, and accurate. The images acquired allow fast and accurate 3D
reconstruction of the sherds with an accuracy within 0.2mm. The system
has been deployed on an archeological site in Armenia this past summer
and demonstrated expected efficacy and robustness.

As the second topic, I will present a novel neural rendering method,
called NeuS, for reconstructing 3D objects from 2D image inputs.
Existing neural surface reconstruction approaches have difficulty in
reconstructing complex objects with severe self-occlusion because they
are prone to get stuck in local minima. Meanwhile, recent neural
rendering methods for novel view synthesis, such as NeRF and its
variants, use volume rendering to achieve more robust optimization, even
for highly complex scenes. However, they cannot extract high-quality
surfaces because of the lack of surface constraints. We introduce
surface constraints into the NeRF framework by representing a surface as
the zero-level set of a signed distance function (SDF) and devise a new
volume rendering method to learn this neural SDF representation.
Extensive experiments show that NeuS outperforms the state-of-the-arts
in high-quality surface reconstruction, especially for objects and
scenes with complex structures and self-occlusion.