Abstract High precision 3D sculpture models provide accurate records of the shape, material, and color of cultural heritage artifacts. This is essential for digital documentation, preservation, and archaeological research. Constructing these models involves geometry modeling and texture reconstruction. However, significant challenges remain in achieving high-precision texture mapping. This paper presents a method for high-precision texture reconstruction based on non-rigid transformation. The process begins with a coarse registration of texture images to the geometrical model using Direct Linear Transformation (DLT). This registration is then optimized using a weighted Thin Plate Spline (TPS) function to minimize local matching errors. Experimental results from the Dunhuang Mogao Grottoes in China demonstrate the efficiency and feasibility of this approach.
Introduction
Creating high-fidelity 3D models of cultural heritage sites requires two primary components: accurate geometry (shape) and accurate texture (color and material). While modern laser scanners are highly effective at capturing geometric data, "wrapping" high-resolution 2D photographs around those 3D shapes—a process known as texture mapping—remains a complex technical challenge. Traditional methods often fail to account for local distortions, leading to visible misalignments in the final model.
The Problem: Rigid Registration Limits
Traditional texture mapping relies on "rigid" registration, which assumes a perfect pinhole camera model. In practice, factors such as lens distortion, laser scanner noise, and data processing errors mean that a simple linear projection often results in noticeable local mismatches. For example, a sculpture's eye or lip texture might be shifted several millimeters away from its corresponding geometric position on the 3D mesh.
Methodology
The researchers propose a two-stage registration pipeline to address these local errors:
1. Coarse Registration (DLT)
First, Direct Linear Transformation (DLT) is used to establish a rough correspondence between 2D image points and 3D model coordinates. DLT is particularly advantageous because it does not require initial camera position values, making it suitable for images captured with standard, non-metric cameras.
2. Fine Non-Rigid Registration (Weighted TPS)
To correct the local errors remaining after the DLT stage, the authors employ a Thin Plate Spline (TPS) function. TPS is a non-rigid coordinator capable of warping the image to fit the geometric model precisely.
The Weighted Improvement: Standard TPS treats all control points as equally accurate. The authors improved this by assigning different weights to control points based on their "confidence gradient." This allows the algorithm to prioritize highly certain points while smoothing over areas where the data might be noisier.
Experimental Results
The method was tested on two real-world datasets from the Dunhuang Mogao Grottoes:
- Bodhisattva (Cave 196): Chosen for its distinct geometric features. While the DLT method showed clear mismatching in fine details like the eye socket, the non-rigid TPS method aligned the texture perfectly.
- Sitting Buddha (Cave 285): This sculpture was severely damaged and had a smooth surface, making it difficult to select control points based on geometry alone. In this case, the researchers used the backscattered intensity (pseudo-color) from the laser scanner to identify features. The weighted TPS method successfully aligned the contours of the chignon (hair bun), which were previously offset under rigid registration.
Conclusion and Future Work
The proposed non-rigid transformation approach effectively overcomes the "bottleneck" of rigid mathematical models in high-precision texture mapping. By accounting for physical realities like lens distortion through weighted TPS, the researchers achieved high accuracy at both global and local scales.
Currently, the process requires manual selection of 2D-3D corresponding control points. Future research will focus on the automation of this selection process, developing algorithms that can automatically identify intrinsic features in both images and 3D models to streamline high-quality texture reconstruction.
Summary updated