3D Reconstruction of Textureless Surfaces

781 concepts585 papers9 proposals

Traditional and modern 3D reconstruction techniques—ranging from classical Multi-View Stereo (MVS) to Neural Radiance Fields (NeRF) and 3D Gaussian Splatting (3DGS)—fundamentally rely on a single core assumption: photometric consistency. This principle states that a physical 3D point should look roughly the same when viewed from different angles. By matching these visual features across multiple images, algorithms can triangulate depth and map out a scene's geometry.

However, as established by Schöps et al. in their work on the ETH3D benchmark, surfaces with minimal texture remain a primary failure mode for state-of-the-art reconstruction pipelines. When an algorithm encounters a large, featureless surface—such as a blank white wall, a polished floor, or a uniform ceiling—photometric consistency breaks down entirely.

Why Textureless Regions Cause "Floaters" and Noise

To understand why textureless regions are so destructive, consider what happens during optimization. In a region lacking distinct visual landmarks, a pixel in one camera view looks identical to hundreds of pixels in an adjacent view. As Xu et al. point out in ACMP, this creates a severe ambiguity where multiple entirely different depth hypotheses appear equally valid to the algorithm.

In neural implicit representations (like Neural SDFs) and point-based rendering (like 3DGS), this manifests as the shape-radiance ambiguity. As Yu et al. demonstrate in MonoSDF, because color-based supervision is the primary signal, the optimization can satisfy the loss function using incorrect geometries. For instance, instead of reconstructing a flat, solid wall, a neural network might generate a cloud of disconnected, semi-transparent primitives in empty space—commonly referred to as "floaters"—or warped, bumpy geometry. As long as these artifacts render as the correct uniform color from the training viewpoints, the network has no mathematical reason to eliminate them. Fu et al. echo this in Geo-Neus, noting that color-based supervision alone provides insufficient signals to resolve smooth or untextured regions accurately.

Constraining the Ambiguity with Geometric Priors

Since the visual data itself is insufficient to resolve the true structure of textureless surfaces, recent research focuses on introducing structural and geometric constraints to "fill in the blanks."

Structural and Planar Assumptions

One powerful intuition is that featureless indoor surfaces are rarely arbitrary; they are typically flat and aligned with the architecture of the room. Xu et al. leverage this by embedding adaptive piecewise planar priors directly into the PatchMatch MVS pipeline, allowing the algorithm to propagate geometric trends from textured edges into the blank interiors of walls. Similarly, Wang et al. introduce StructNeRF, which enforces patch-based consistency and planar constraints to resolve the shape-radiance ambiguity in indoor scenes using sparse views.

Monocular Depth and Normal Priors

Another dominant strategy is to supplement multi-view images with monocular geometric cues derived from neural networks trained on massive datasets. Rather than relying solely on cross-view matching, frameworks like MonoSDF and DN-Splatter (developed by Turkulainen et al.) force the 3D representation to respect predicted depth maps and surface normals. By constraining the orientation and position of the underlying primitives (such as Gaussians or implicit surfaces) to match these smooth geometric priors, the optimization successfully eliminates floaters and prevents the formation of noisy, disconnected geometry.

Coarse-to-Fine Guidance

Alternatively, the ambiguity can be managed by leveraging multi-scale geometry. Wang et al. propose MG-MVS, a pyramid architecture that extracts a coarse 3D surface mesh where large-scale geometric trends are easier to capture. This coarse mesh is then used to guide and constrain depth estimation at finer, high-resolution scales, preventing the pixel-matching process from getting lost in uniform, low-texture regions.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 558 papers
  2. Version 1 · 2026-10-04 · 605 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_textureless_surface_reconstruction_1a64cfde,
  title = {textureless surface reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/textureless-surface-reconstruction/rdr_1a64cfdee1b04f9f9066d0f4b0e1bd2a},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_textureless_surface_reconstruction_1a64cfde_v2,
  title = {textureless surface reconstruction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/textureless-surface-reconstruction/rdr_1a64cfdee1b04f9f9066d0f4b0e1bd2a/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake