Making bas-reliefs from photographs of human faces

Abstract

Bas-reliefs are a form of flattened artwork, part-way between 3D sculpture and 2D painting. Recent research has considered automatic bas-relief generation from 3D scenes. However, little work has addressed the generation of bas-reliefs from 2D images. In this paper, we propose a method to automatically generate bas-relief surfaces from frontal photographs of human faces, with potential applications to e.g. coinage and commemorative medals.Our method has two steps. Starting from a photograph of a human face, we first generate a plausible image of a bas-relief of the same face. Secondly, we apply shape-from-shading to this generated bas-relief image to determine the 3D shape of the final bas-relief. To model the mapping from an input photograph to the image of a corresponding bas-relief, we use a feedforward network. The training data comprises images generated from an input 3D model of a face, and images generated from a corresponding bas-relief; the latter is produced by an existing 3D model-to-bas-relief algorithm. A saliency map of the face controls both model building, and bas-relief generation.Our experimental results demonstrate that the generated bas-relief surfaces are smooth and plausible, with correct global geometric nature, the latter giving them a stable appearance under changes of viewing direction and illumination.

Use of non-photorealistic rendering and photometric stereo in making bas-reliefs from photographs

“…The majority of recent bas-relief generation methods [1,2,3,4] start with a depth-map of a 3D scene, and selectively compress depths to create a bas-relief surface. Relatively fewer works [5,6,7] have considered using 2D images as input.…”

“…In summary, we show how NPR can be combined into our previous method [7] and how photometric stereo can be used instead of SFS. The resulting contributions of this paper are as follows.…”

“…Manual production of bas-reliefs requires considerable artistic skill, and is also time consuming. Recent research in computer aided design has considered automatic bas-relief generation from 3D scenes [1,2,3,4] and 2D images [5,6,7]. As images are much cheaper and easier to capture than 3D models, starting from the former potentially has much wider application.…”

“…An alternative approach is to use photometric stereo [10], which requires two or more images taken under different (known) lighting directions, from a single fixed viewpoint. In our previous work [7], we relight the image to a standard direction (instead of estimating the original lighting). The same image relighting technique also enables us to produce multiple images, allowing the use of photometric stereo instead of SFS for shape recovery.…”

“…In our previous work [7], we reported a method to generate bas-reliefs from frontal human face photographs using the technique of shape from shading (SFS). Our two-step pipeline combined techniques from neural networks, image relighting, and SFS.…”