Automatic inference of a anatomically meaningful solid wood texture from a single photograph

TL;DR

The paper tackles automatic appearance matching of solid wood textures from a single photograph, aiming to recover anatomically meaningful internal structure (growth rings) for volumetric wood rendering. It introduces curved Gabor filters to detect rings and a phase-based loss to recover ring deformations, enabling automatic initialization for a procedural wood model based on Liu et al. 2016, expressed through a loss $E = (J_{\boldsymbol{\u03c6}} - I_{\boldsymbol{\u03c6}})^2$ that aligns ring phase. The contributions include automatic estimation of board orientation, a radial distortion field, and a color map to instantiate a 3D solid wood texture, plus demonstrations of volumetric rendering and dendrochronology applicability. The approach yields realistic solid wood textures from photographs with meaningful internal structure, enabling applications in furniture fabrication, translucent wood renderings, and data-rich dendrochronology analyses; future work points to improved subsurface scattering models (e.g., BSSRDF) and faster approximate rendering.

Abstract

Wood is a volumetric material with a very large appearance gamut that is further enlarged by numerous finishing techniques. Computer graphics has made considerable progress in creating sophisticated and flexible appearance models that allow convincing renderings of wooden materials. However, these do not yet allow fully automatic appearance matching to a concrete exemplar piece of wood, and have to be fine-tuned by hand. More general appearance matching strategies are incapable of reconstructing anatomically meaningful volumetric information. This is essential for applications where the internal structure of wood is significant, such as non-planar furniture parts machined from a solid block of wood, translucent appearance of thin wooden layers, or in the field of dendrochronology. In this paper, we provide the two key ingredients for automatic matching of a procedural wood appearance model to exemplar photographs: a good initialization, built on detecting and modelling the ring structure, and a phase-based loss function that allows to accurately recover growth ring deformations and gives anatomically meaningful results. Our ring-detection technique is based on curved Gabor filters, and robustly works for a considerable range of wood types.

Figures (6)

• Figure 3: Ring detection process. We start with an input 2D texture (left) and compute its local orientation. We then compute a curved region around each point in the image and apply a Gabor filter on it. Accumulating over all pixels gives the Gabor filter phase response. We then trace the rings in the phase image to detect their positions (right).
• Figure 4: The procedural wood model is an infinite 3D texture centered around a virtual tree (left). An arbitrary 3D mesh, such as a cube (middle), can be positioned within the 3D texture coordinate system accordingly to where the object was carved from the tree. This gives a textured solid wood object (right) that can be rendered.
• Figure 5: Volumetric appearance is highly dependent on grain orientation - shown here are renderings of 1.8mm thick veneer sheets backlit by a spherical light source. The volumetric properties are derived from our fit (right), and extruded along the surface normal (left). Insets show side-views to illustrate grain
• Figure 6: Volumetric rendering of a lampshade that was cut from a fit model using our method (Fig. \ref{['fig:results']}, 2nd column). Both single scattering albedo and volumetric density are heterogeneous.
• Figure 7: Results obtained with our approach. Rows from top to bottom: Input texture, detected rings, frontal view of fit model (diffuse reflectance only), volumetric visualization (false color), fit color map.