In contemporary society, we are surrounded by not only physical materials, but also images of them. We are capable of judging materials and their properties with only visual information. For instance, if an object looks solid and glossy or soft and fluffy. This ability is called
...
In contemporary society, we are surrounded by not only physical materials, but also images of them. We are capable of judging materials and their properties with only visual information. For instance, if an object looks solid and glossy or soft and fluffy. This ability is called material perception. As for images, there are various ways of image making, such as photography, painting, computer rendering, etc. And a new method has emerged recently: generative AI. All these image generation methods can produce different appearances of the same object or material. In this thesis, we studied human visual perception of two types of appearances: appearance as material property, appearance as pictorial style and the interaction between them.
In Chapter 2, we investigated depiction style by zooming in on a single motif, an apple. By using the fragments instead of the whole painting, we were able to keep the subject matter relatively constant, and isolate style from composition as well as other contextual information. We first constructed a perceptual space of style using similarity judgements from online participants. Then we fitted perceived attributes to this space to understand its dimensions. The data resulted in a three-dimensional space. Dimension 1 is associated with smoothness and brushstroke coarseness. Dimensions 2 and 3 are related to hue and chroma. Surprisingly, we also found a rotational relation between creation year and the first two dimensions, revealing a certain cyclic, repetitive pattern of style. The results suggest style can already be perceived in fragments of paintings.
In Chapter 3, we studied the influence of medium on appearance. For example, imagine an oil-painted apple and a pencil-sketched apple: they can have different appearances. The comparison between different media has rarely been studied. One possible reason is the difficulty to isolate medium from its confounding factor, subject matter. We found a solution by comparing oil paintings and their engraved reproductions. The identical content gave us a perfect opportunity to compare material perception from two distinct media. We collected 15 pairs, consisting of 88 fragments depicting different materials like fabric, skin, wood and metal. We also created three manipulations to understand the effect of color (a grayscale version) and contrast (equalized histograms towards both painting and engraving). We collected ratings on five attributes: three-dimensionality, glossiness, convincingness, smoothness and softness. Paintings showed a broader perceived range than engravings, with contrast equalization having a greater impact on perception than color removal. Possibly engravers used local contrast to compensate the absence of color.
In Chapter 4, we analyzed an emerging medium from a non-human creator, generative AI. In two experiments, we explored human material perception using generative AI stimuli and compared the perceptual spaces of three generative AI models, as well as a computer-generated BRDF stimulus set, the MERL dataset. In Experiment 1, we used text descriptions of 32 materials from MERL (e.g. ‘green fabric’) as prompts for DALL-E 2 and Midjourney v2. Both AI models resulted in a 2D space while MERL resulted in a 1D one. The three spaces showed low similarity, suggesting the AI models generated unique and different images of materials from identical text prompts. In Experiment 2, we explored another text-to-image model Stable Diffusion v1.5 with an add-on, ControlNet. ControlNet allowed us to add additional graphical constraints besides text input. In this way we could inspect more complex shapes. We kept the same 32 descriptions and generated material blobs in three shapes, from simple to more complex geometry. The three perceptual spaces from the three shapes showed high similarity, indicating both robust structure and minor influence of object shape on material perception. Interestingly, the perceptual spaces from Experiment 2 also shared similar structure as perceptual spaces from other material studies using real-world photos, computer renderings and depictions. In sum, we investigated visual perception through the lens of art by examining appearances rendered by painters, engravers and generative AIs.@en