Bas-relief generation controlled by different height maps

Jing Shang, Wen-Jie Yu, Meili WangJournal of Image and Graphics 2020 · DOI

This paper, published in the Journal of Image and Graphics (2020), presents a refined approach to digital bas-relief generation—the art of creating 3D sculptures that project minimally from a flat background.

TL;DR

The researchers propose a method to generate stylized digital bas-reliefs by using 2D "masks" to precisely control the height and detail of different regions. By combining mask theory with advanced normal filtering (GRNF) and surface reconstruction (SfG), the system allows designers to create complex, multi-layered decorative patterns for industrial design and home decor with high artistic flexibility.

What Problem Does This Paper Address?

Traditional bas-relief carving is a labor-intensive craft that is difficult to modify once started. While digital methods exist—such as Shape-from-Shading (SfS) or compressing 3D models into thin volumes—they often struggle with "style diversity."

Most automated tools apply a uniform transformation to an entire image or model. This makes it difficult for a designer to say, "I want this part of the relief to be deep and sharp, but that part to be shallow and smooth." Furthermore, existing methods often produce artifacts at depth discontinuities (edges where the height jumps abruptly). The authors sought a way to give designers localized, intuitive control over the relief's topography.

Methods

The authors introduce a pipeline that bridges 2D image processing and 3D geometry:

  1. Height Map Acquisition via Masking: This is the core innovation. Instead of a simple global height assignment, the user applies different types of masks (binary, grayscale, or textured) to an input image. These masks act as "height controllers," allowing specific regions to be elevated or recessed independently.
  2. Normal Decomposition (GRNF): The system uses Gaussian Mixture Model-based Rolling Guidance Normal Filtering. This algorithm decomposes the surface's "normals" (the directions the surface points) into a structural base layer and a fine detail layer. This separation allows the system to preserve sharp edges while smoothing out unwanted noise.
  3. Surface Reconstruction (SfG): To turn these processed normals back into a 3D surface, the authors use Surface from Gradients (SfG). This involves a two-step optimization: a local adjustment to shape the mesh and a global adjustment (using least-squares) to ensure the entire surface is mathematically consistent and continuous.
  4. Post-processing: Laplacian and bilateral filters are applied to the final mesh to ensure smooth boundaries and eliminate any remaining "stair-step" artifacts.

Datasets

As this is a procedural algorithmic paper rather than a deep learning one, it does not rely on a massive training dataset. Instead, the authors evaluated their method using:

  • Standard 3D Benchmarks: The "Stanford Bunny" was used to demonstrate denoising and smoothing capabilities.
  • Custom 2D Inputs: Various decorative patterns, including cat silhouettes for pillows and traditional Chinese motifs for furniture (vases, tables), were used to showcase the masking and height control features.

Key Insights and Ideas

  • 2D Masks for 3D Control: The "aha moment" is treating 2D masks not just as visibility toggles, but as depth-control interfaces. This allows a designer to "paint" the desired thickness of a relief using familiar 2D tools.
  • The Three Knobs of Relief: The authors identify three critical parameters that define the "look" of a relief:
  • $\mu$ (Detail Richness): Higher values act like a sharpening filter, making fine textures pop.
  • $\beta$ (Flatness): Controls the overall "smoothness" or "planarity" of the relief.
  • $h$ (Height Factor): Directly scales the physical thickness of the resulting model.
  • Normal Domain Processing: By manipulating the surface normals rather than the raw coordinates, the method avoids the jagged edges and distortions common in simpler height-map-to-3D conversions.

Key Results

  • Localized Stylization: The researchers successfully demonstrated that they could generate a single relief where different parts have different visual "weights"—for example, a cat pattern on a pillow where the cat's body is smooth but its features are sharp and distinct.
  • Industrial Application: The paper showcases practical applications in the furniture industry, generating relief-decorated table legs and vases that are ready for 3D printing or CNC milling.
  • Visual Quality: Compared to traditional global compression methods, this approach maintains clearer outlines and richer details, even when the total depth of the relief is very shallow.

Limitations and Future Work

The authors acknowledge a few areas for improvement:

  • Single-Model Focus: The current method is optimized for controlling the height of a single relief model at a time.
  • Style Transfer: A promising future direction is the integration of "style transfer" algorithms, which would allow the system to automatically apply the artistic texture of one relief (e.g., a "stone-carved" look) to the content of another.
  • Automation: While the masking provides great control, it still requires manual input. Future work could involve automated mask generation based on image saliency or semantic segmentation.

Summary updated

Abstract

目的数字浅浮雕是在平面载体上塑造高低起伏形象的一种造型艺术,具有独特的结构及视觉效果,应用场景极为广泛。为了增加数字浮雕设计的多样性,实现浮雕的风格化应用,提出一种基于高度图的浅浮雕模型生成方法。方法引入掩模理论,利用掩模操作对待处理的图像进行处理,融合图像处理技术,控制所要生成浮雕不同部位的高度,得到控制浮雕生成效果的高度图,借鉴已有的浅浮雕模型生成方法,利用基于高斯混合模型的滚动引导法向滤波(Gaussian mixture model based rolling guidance normal filtering,GRNF)进行法向分解,基于SfG(surface from gradients)的局部调整和全局重建的方法进行曲面重建,采用拉普拉斯算子及双边滤波器进行去噪平滑处理,最终生成不同高度风格的浅浮雕模型。结果实验结果表明,本文方法能够生成带有不同视觉效果的浅浮雕模型,通过对细节特征及结构特征相关参数的调整,均能够生成轮廓清晰、细节丰富的浅浮雕模型。结论本文提出的基于高度图的浅浮雕模型生成方法能够为浅浮雕的多样化设计提供新的思路和方法,在家居装饰行业中有重要的应用价值。
BibTeX
@article{shang2020bas,
  author = {Jing Shang and Wen-Jie Yu and Meili Wang},
  title = {{Bas-relief generation controlled by different height maps}},
  journal = {Journal of Image and Graphics},
  year = {2020},
  doi = {10.11834/jig.190630},
}
Figure