Advancements in 3D Reconstruction and Scene Understanding

The recent developments in the field of 3D reconstruction and scene understanding have been significantly influenced by advancements in neural implicit representations and radiance fields. A notable trend is the focus on overcoming challenges related to limited image overlap, reflective surfaces, and the reconstruction of complex geometries in both macro and micro scales. Innovations include hybrid methods that leverage cross-representation uncertainty estimation for accurate reconstruction in urban environments, techniques for handling mirroring surfaces without prior annotations, and approaches for single-view reconstruction of volumetric fields with multiple light scattering effects. Additionally, there's a growing emphasis on scalable and high-quality neural implicit representations that enable detailed and accurate 3D reconstructions by modeling scenes as a fusion of multiple independent local neural SDFs. Another significant advancement is the integration of multi-modal data, such as combining White Light Interferometry with Optical Microscopy, to achieve detailed 3D reconstructions with natural color textures for microscale research applications.

Noteworthy Papers

  • CoStruction: Introduces a hybrid implicit surface reconstruction method for urban scenes with limited camera overlap, outperforming state-of-the-art methods.
  • NeRFs are Mirror Detectors: Presents a method for reconstructing scenes with mirroring surfaces without prior annotations, leveraging photometric inconsistencies.
  • Light Transport-aware Diffusion Posterior Sampling: Offers a novel approach for single-view reconstruction of volumetric fields, significantly improving quality.
  • Scalable and High-Quality Neural Implicit Representation: Proposes a versatile method for 3D reconstruction by fusing multiple local neural SDFs, enabling high-fidelity and scalable scene reconstruction.
  • OpticFusion: Introduces a multi-modal approach for 3D reconstruction of microstructures, combining WLI and OM images to achieve detailed reconstructions with natural color textures.

Sources

CoStruction: Conjoint radiance field optimization for urban scene reconStruction with limited image overlap

A continuous scale space of diffeomorphisms

NeRFs are Mirror Detectors: Using Structural Similarity for Multi-View Mirror Scene Reconstruction with 3D Surface Primitives

Light Transport-aware Diffusion Posterior Sampling for Single-View Reconstruction of 3D Volumes

Representation Learning of Point Cloud Upsampling in Global and Local Inputs

Matching Free Depth Recovery from Structured Light

Scalable and High-Quality Neural Implicit Representation for 3D Reconstruction

OpticFusion: Multi-Modal Neural Implicit 3D Reconstruction of Microstructures by Fusing White Light Interferometry and Optical Microscopy

Normal-NeRF: Ambiguity-Robust Normal Estimation for Highly Reflective Scenes

Built with on top of