Talks and Poster Presentations

Fidelity Loss In Wiki Art Images Of Paintings: Effects On Aesthetic Ratings And Preferences

September 11, 2025

Talk, Sixteenth International Conference on The Image, Paris, France

With the emergence of large online datasets, research with digital versions of images is more popular than ever. However, little work has been done to assess the fidelity of these digital reproductions. We identify four types of fidelity loss in the WikiArt dataset and assess their effect on aesthetic judgement and originality assessment in two online studies. Our Fidelity Of Painting Images Dataset (FOPID) contains 750 images spanning 25 art styles, including 250 low-fidelity images sampled from WikiArt (blurred, cropped, enhanced, outdated, or combinations thereof), and two higher-fidelity counterparts, one from a museum website and one from a third-party source (with in-between quality). In Study 1, lay participants (N=298) rated 100 of the 250 images (randomly in one of the three versions) on their aesthetic appreciation (0–100 scale). No significant difference was found between fidelity groups , although images with combined distortions received the lowest mean rating. In Study 2, lay participants (N=108) saw the three variants of each painting side by side, and indicated their preferred image and the image they believed the original. The in-between version was most preferred, but the high-fidelity version was most often identified as the original. Overall, fidelity appears to matter little without context, although different distortions lead to different aesthetic evaluations. When different versions can be compared directly, participants prefer higher-fidelity images, calling into question the generalizability of studies relying on low- fidelity reproductions and underscoring the importance of fidelity awareness in aesthetics research. Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925, GRAPPA) awarded to JW.

Evaluating the WikiArt Dataset using Quantitative Image Properties, Ratings, and Art History

August 25, 2025

Poster, 47th European Conference on Visual Perception (ECVP), Mainz, Germany

The WikiArt dataset is popular in research on art and aesthetics. However, closer inspection reveals occasional subpar image quality, possibly distorting the results of previous research. Here, we present a novel dataset on image quality and findings from a preregistered online study to investigate aesthetic appreciation in relation to common image distortions in WikiArt. Our stimulus set (750 images, 25 art styles) was curated by an art historian, who identified four quality issues within the WikiArt dataset. Many images are either blurred, cropped, enhanced, outdated or a combination of these factors. All 250 (50 per issue) low-quality images from WikiArt were labeled and matched with two high-quality counterparts. We used 38 quantitative image properties (QIPs) to specify the four quality issues labelled by the expert. A one-way MANOVA revealed a strong multivariate effect (V = 0.935, p<.001) of label membership on the QIPs. To reduce dimensionality, we performed PCA and FA and found that the first 5 to 7 components captured ~58% to 60% of total variance, respectively. To investigate the effect of image quality (low vs. high) on aesthetic appreciation, each participant in an online study (N=332) rated 100 images with a Slider Rating Scale (0-100). We also collected demographic information, familiarity with the artworks as well as art expertise. A Mixed Effects Bayesian Regression indicates image quality had only a small, nonsignificant negative effect on the ratings (Est. -1.33, 95%CI[-3.17, 0.53]), even including expertise (Est.-4.13, 95%CI[-10.90, 2.53]) and its interaction with quality (Est. 1.04, 95%CI[-0.92, 2.95]). We provide a theoretical and quantitatively validated dataset of quality distortions in artworks in WikiArt. Model results suggest the quality issues do not influence aesthetic appreciation in a one- image-at-a-time rating task. For deeper insights a follow-up study of direct image comparisons across quality versions is currently being conducted. Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925, GRAPPA) awarded to JW.

DODA: Introducing A Database of Datasets for Aesthetics

August 23, 2025

Poster, Visual Science of Art Conference (VSAC), Wiesbaden, Germany

Stimulus selection is a critical first step in the empirical and computational investigation of vision and art. While many image and stimulus sets have been curated and published alongside valuable data, it can be difficult to find the best set for one’s own research, as they often differ in various aspects and are shared across different platforms. Therefore, we present a Database of Datasets for Aesthetics (DODA), an intuitive Web Application in which researchers can browse all important datasets for aesthetics research, with annotated properties that are highly relevant for their possible suitability. Based on a systematic review of stimulus sets in aesthetics, DODA offers a centralized search system, intended to promote open-science oriented research practices. Currently, DODA encompasses over 50 datasets, with a link to their location on the web. We compare the datasets and document relevant selection criteria, such as Size, Number of Participants, Participants per Stimulus, Research Question, Rating Scales, Stimulus Source, Resolution, Homogeneity, etc. The information is displayed in a table for users to easily filter through, prioritizing the criteria most important to them. DODA also provides a wide range of quantitative image properties (QIPs) computed for all datasets for users to download. Reusing a dataset is not only ecological for the researcher and environment, but it also allows for the comparison of different variables across the same stimulus set, increasing our collective understanding of aesthetic appreciation. Working on the same datasets provides a feasible opportunity to directly and indirectly collaborate across methodologies within aesthetics. For DODA to grow alongside the field of aesthetics, we encourage researchers to suggest other aesthetics datasets to be included in the future. DODA is part of the Aesthetics Toolbox, which encompasses useful tools like image resizing and calculating QIPs to make aesthetics research on digital stimuli more accessible and reproducible Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925) awarded to JW.

Composition and Spatial Layout: A Comparative Investigation of the two concepts in Images of Natural Scenes and Artworks

June 04, 2025

Poster, Lentecyclus, Koninklijke Vlaamse Academie van België voor Wetenschappen en Kunsten (KVAB), Brussels, Belgium

While Composition is a term commonly used in art-theoretical discussions, and art education, Spatial Layout (hereafter “Layout”), is more common in the literature on depth and scene perception. With two large online studies (N= 1364 and N= 1308) of 160 artworks and 160 natural scenes, respectively, we aim to provide a foundation for future work by investigating the relationship of the two concepts as well as important aesthetic measures, namely Pleasure, Interest, Order, and Complexity. In both studies, our participants were randomly assigned to either the Composition or the Layout condition and received definitions and examples of good and poor Composition or clear and unclear Layout. They then viewed 40-50 randomly selected images in two blocks. In the first block they rated all images on their Composition or Layout and either Pleasure or Interest, and either Order or Complexity, using 7-point Likert Scales. In the second block they rated the same images again (new random order) on Composition or Layout and on the two remaining aesthetic concepts. Participants also filled out a standard demographics questionnaire, an art-experience questionnaire and scales from selected personality questionnaires for Openness to Experience, Need for Closure, Sensation Seeking and Aesthetic Sensitivity. Within artworks Composition and Layout are positively correlated with Order (rs = 0.91, rs =0.84), Pleasure (rs = 0.87, rs =0.86), Interest (rs = 0.58, rs =0.64) and negatively correlated with Complexity (rs = -0.32, rs = -0.22). Within images of natural scenes, these findings are similar (Order (rs = 0.66, rs = 0.74), Pleasure (rs = 0.77, rs = 0.54), Interest (rs = 0.70, rs = 0.43) and negatively correlated with Complexity (rs = 0.44, rs = -0.12)), albeit the correlation coefficients are smaller. In artworks as well as natural scenes, Composition and Layout were judged reliably and their Spearman’s correlation coefficients are 0.90 and 0.74, respectively. Composition and Layout are correlated higher for representative artworks than for abstract artworks. Our results indicate that Composition and Layout are closely related concepts but are not completely overlapping, as their relationship appears to be at least partly dependent on the image category, and their respective relationships with aesthetic determinants and measures vary. Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925) awarded to JW.

Symmetry Detection and Figure-Ground Segmentation in Perceptual Organization of Rich Images

May 19, 2025

Poster, 25th Annual Meeting of the Vision Science Society (VSS), St Pete Beach, FL, USA

Symmetry plays a critical role in perceptual organization, contributing to tasks like perceptual grouping, figure-ground segmentation, shape perception and aesthetics. However, most studies rely on simple geometric patterns, leaving its role in complex, real-world stimuli underexplored. Here, we present findings from a preregistered online study involving 901 participants and a set of 200 real-world photographs and 200 artworks to investigate symmetry detection and its interplay with figure-ground segmentation. Participants identified symmetric regions in images by superimposing bounding boxes and axes of symmetry on self-selected image regions, allowing multiple regions per image. For each region, they rated salience (how much it stood out to them) and strength of symmetry (ranging from imperfect to nearly perfect). These human annotations were compared to symmetry detection by a deep learning model (U-Net). Additionally, object detection (YOLO11) and segmentation (SAM2) algorithms were used to explore how figure-ground segmentation influenced symmetry perception. Our results confirm the dominance of vertical symmetry axes in real-world scenes (72.91%) and artworks (73.87%). Bayesian regression analysis revealed that as the number of marked regions increased, their salience and size decreased while symmetry strength slightly increased. The U-Net model showed alignment with human annotations in high-salience, high-symmetry regions, with overlaps of 55% and 53% for natural scenes and artworks, respectively. Overall performance was better for natural scenes (Weighted F1-Score = 0.43) than for artworks (WF1 = 0.40), but it is clear that machine vision models still have a way to go before they can mimic human symmetry perception. Critically, symmetry detection was heavily influenced by object presence and recognizability within the image, suggesting a strong interaction between symmetry perception, figure-ground segmentation, and object detection. These findings bridge the gap between traditional symmetry research and real-world visual contexts, offering insights into the computational and perceptual mechanisms underlying symmetry detection and in complex images.

Spatial Layout and Composition: Related but Distinct Factors for the Aesthetic Appreciation of Natural Images

August 28, 2024

Poster, 46th European Conference on Visual Perception (ECVP) 2024, Aberdeen, Scotland

Spatial Layout is an important concept in the literature on space and scene perception, while Composition is central in theoretical aesthetics, and art history, philosophy, and education. Both factors depend strongly on perceptual organization, and they could therefore have a similar impact on aesthetic appreciation. We aimed to reveal similarities and differences between both concepts in relation to aesthetic appreciation, by conducting an extensive online study (N=1300) with a diverse stimulus set of real-world images. We collected 7-point ratings for Spatial Layout, Composition, Order, Complexity, Pleasure, and Interest on 160 images of 80 manmade and 80 natural scenes. Participants rated either Composition or Spatial Layout of 40 images in two blocks, after they received a brief explanation with examples of good and bad composition, or clear and unclear spatial layout. In the first block they also rated either Pleasure or Interest, and either Order or Complexity. In the second block they rated the same images again on Composition or Spatial Layout and the remaining two concepts. Participants also completed questionnaires for basic demographics, art-experience, and personality. First analyses with Spearman correlations confirmed that Spatial Layout and Composition can be judged reliably (0.73 and 0.77, respectively) and were highly correlated (0.74), with some unexplained variance suggesting that they are not completely overlapping concepts. Composition was more relevant for Pleasure and Interest (0.77 and 0.70, respectively) than Spatial Layout (0.54 and 0.43, respectively). Order correlated more strongly with Spatial Layout (0.74) than with Composition (0.66), while for Complexity the correlations were both weak but had opposite signs (0.26 for Composition, -0.12 for Spatial Layout). Our findings indicate that Spatial Layout and Composition are related but distinct factors, with unique relationships to different dimensions of aesthetic appreciation. Future analyses of this dataset will provide insight into moderators of these relationships.

Composition and Spatial Layout of Images of Artworks in relation to their Aesthetic Appreciation

May 09, 2024

Poster, International Association of Emprical Aesthetics (IAEA), Palma de Mallorca, Spain

While Composition has been at the center of philosophical aesthetics, art-theoretical discussions, and art education, it has received considerably less attention in empirical aesthetics. The same is true for the notion of Spatial Layout (hereafter “Layout”), which is more central in the literature on space and scene perception. With this large preregistered (https://osf.io/67mx5) online study of 160 artworks, we aim to provide a foundation for future work by investigating the relationship of important aesthetic measures, namely Pleasure, Interest, Order, and Complexity, with Composition and Layout. Our participants were randomly assigned either the Composition or the Layout condition, received definitions and examples of good and poor Composition or clear and unclear Layout, and then viewed 50 randomly selected artworks in two blocks. In the first block they rated them on either Pleasure or Interest, either Order or Complexity, and either Composition or Layout, using 7-point Likert Scales. In the second block they rated the same images again (new random order) on Composition or Layout and on the two remaining aesthetic concepts. Participants also filled out a standard demographics questionnaire, an art-experience questionnaire and scales from selected personality questionnaires for Openness to Experience, Need for Closure, Sensation Seeking and Aesthetic Sensitivity. First results (N=494) show that high ratings for Composition and Layout lead to higher Pleasure ratings. Ratings for Order are highly positively correlated with ratings for Composition and Layout, while ratings for Complexity are negatively correlated with ratings for Composition and Layout. Participants scoring high on Openness and Sensation Seeking require lower scores on Composition and Layout to give higher Pleasure ratings. Composition and Layout are correlated higher for representative artworks than for abstract artworks, and their relationship with other concepts seems to be slightly different. Data collection is still ongoing until we have ratings from 1280 participants. Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925) awarded to JW.

Symmetry Detection in Images of Natural Scenes by Humans and Machines

August 31, 2023

Poster, 45th European Conference on Visual Perception (ECVP) 2023, Paphos, Cyprus

Although mirror symmetry is an established and popular principle of perceptual organization, human symmetry detection in images of natural scenes remains highly understudied, when compared to symmetry detection in artificially created dot patterns and shapes. In this multidisciplinary project, we investigate human symmetry detection in 100 images of natural scenes in relation to quantitative metrics derived from computer vision and machine learning. In our study participants were asked to place a rectangular bounding box around an image region they perceived as mirror-symmetric and to indicate the axis of symmetry. They could place as many bounding boxes as they saw fit. For each of them, they also rated the perceived saliency of the region (i.e., how much it popped-out from the background) and the strength of the symmetry (i.e., from rather imperfect to almost perfect symmetry). Statistical analysis of 2173 symmetries by 17 participants so far reveals that participants selected bigger, more salient regions of symmetry first. Vertical axes were much more frequent (around 75%) than horizontal and oblique ones. Horizontally and vertically symmetric regions were found to be more salient and more symmetric than oblique ones. Saliency and strength ratings were moderately correlated (around 0.4) across all regions and images. We used different metrics for image quality assessment to compute symmetry accuracy scores for the bounding boxes, revealing large discrepancies between human and computational symmetry assessment (correlations below 0.1), both for saliency and strength. This emphasizes the need to go beyond traditional computer vision algorithms and employ deep learning models. Human data collection is still ongoing, and we also plan to train a deep learning model on symmetry detection and present it alongside these findings. Open data and methods: https://osf.io/9tf4e/ Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925) awarded to JW.

Saliency and strength of mirror symmetry in images of artworks in relation to appreciation and computational metrics

August 26, 2023

Poster, Visual Science of Art Conference (VSAC), Nicosia, Cyprus

The relationship between mirror symmetry and aesthetic appreciation has intrigued vision scientists, empirical esthetics researchers and artists alike, but concrete evidence remains somewhat elusive to this day. In this multidisciplinary project, we investigate human symmetry detection for 100 images of artworks and relate these behavioral data to aesthetic appreciation and computational metrics. Participants were asked to place a rectangular bounding box around an image region they perceived as mirror-symmetric and to indicate the axis of symmetry. They could place as many boxes as they saw fit. For each of them, they also rated the perceived saliency of the region (i.e., how much it popped-out from the background) and the strength of the symmetry (i.e., from rather imperfect to almost perfect symmetry). Statistical analysis of 2839 symmetries by 23 participants so far reveals that participants selected bigger regions of symmetry first and rated them higher on salience and strength of symmetry. Vertical axes of symmetry were most frequently indicated (around 80%). We used different metrics for image quality assessment to compute symmetry accuracy scores for the bounding boxes, revealing large discrepancies between participant ratings and objectively computed symmetry strength. These discrepancies between human and computational symmetry assessment emphasize the need to go beyond computer vision and employ deep learning models. Aesthetic liking of the images, rated by a different pool of observers, seems to be independent from both strength and saliency ratings (correlations <.1). This could be because mirror symmetry is only one aspect of good composition. Human data collection is still ongoing, including aesthetic judgements from the same participants. Additionally, we will train a deep learning model on symmetry detection and figure-ground segmentation, which we will present alongside these findings. Open data and methods: https://osf.io/9tf4e/ Acknowledgment: This work is funded by an ERC Advanced Grant (No. 101053925) awarded to JW.

How to form perceptual memory of multistability: Bias perception but gently

August 29, 2022

Talk, 44th European Conference on Visual Perception (ECVP) 2022, Nijmegen, The Netherlands

When multistable displays are presented intermittently with a long blank interval, they become stabilized via perceptual memory. However, we still lack an understanding of perceptual memory’s role in daily vision, its mechanisms, and even the conditions that lead to its formation. Therefore, we used a reverse correlation method to recover a biasing sequence that forms perceptual memory. In the experiment, participants reported on the direction of rotation of intermittently presented (800 ms on, 1000 ms off) kinetic-depth effect displays. We interleaved fully ambiguous probes (to read out current state of perceptual memory) with biased primes (size/distance cues), randomly disambiguated during the first 300 ms with 10 different bias levels, each lasting 30 ms. We collected 10,000+ trials for three participants. First, we validated the method by computing an average sequence that produced an opposite perceptual dominance in primes: A moderate (50%) bias in favor of the suppressed percept that is gradually reduced to full ambiguity. Next, we repeated the analysis but computed average bias sequences that preceded the dominance change in the following probe. I.e., a prime biasing sequence that established a new perceptual memory. We found the same pattern but with a weaker initial bias. We also computed a bias sequence that reversed the probe (formed perceptual memory) but not the prime itself. Here, an even gentler bias was required. Additional data is necessary to confirm its consistency as such events were rare. In short, to form perceptual memory: gently bias onset perception itself. Open data and methods: https://osf.io/gw7kc/

Proximity, expectations, and attention but not necessarily physics determine perception for simultaneously viewed multiple bistable displays.

August 23, 2021

Poster, 43rd European Conference on Visual Perception (ECVP) , Online

When several multistable displays are viewed at the same time, typically, the same perceptual state tends to be dominant for all stimuli, and perceptual switches tend to occur at the same time (so-called perceptual coupling). We investigated whether this can be altered by an opportunity for a physical interaction between objects. We used a well established version of the walker-on-a-ball display plus a novel display consisting of two rotating gears. In the default configuration, perception for both displays was congruent with physically interacting objects. We gradually altered displays to produce either an abrupt change to the potential interaction (e.g., moving objects away from each other) or to keep it constant despite the visual changes (disambiguating one of the objects). We fit four models that assumed 1) independence of perception of the stimuli, 2) dependence on the stimulus’s properties, 3) dependence on physical configuration alone, and 4) an interaction between stimulus properties and a physical configuration. For the gears display, the perception depended on the stimulus properties, as in perceptual coupling, rather than on the possibility of physical interaction. Regarding the walker-on-the-ball, the perception depended neither on stimulus nor on the possibility of physical interaction but on whether participants were asked to respond on the relative motion of both objects or the absolute motion of the walker alone. This suggests that perception of the walker-on-a-ball was driven primarily by expectations. The results reveal multiple perceptual mechanisms acting at various levels of processing, whereas priors of physical interaction had little influence.

Perceptual Dominance in Face Rivalry Is Driven by Low-Level Properties

August 27, 2019

Poster, 42nd European Conference on Visual Perception (ECVP) 2019, Leuven, Belgium

Two superimposed semitransparent orthogonally oriented faces produce perceptual rivalry with one face being clearly perceived at a time and perception continuously switching between them. We investigated whether perceptual dominance of an individual face is determined by high-level properties, such as gender, age, or emotion, or low-level properties. To this end, we used 20 female and 20 male faces, aged 20 to 25 years, from the Chicago Face Database. They were randomly paired using a round-robin tournament schedule (eight blocks, 20 trials each). Participants viewed a face pair and continuously indicated which face they currently perceive via key presses. We computed two measures of face dominance, (a) as a proportion of trials in which it was the first face perceived at onset and (b) as a proportion of time it was dominant throughout the trial. An exploratory data analysis using linear mixed models showed no systematic relationship between either of the two measures and high-level face descriptors, such as gender, age, or emotions (see https://osf.io/q2fjd). We conclude that in face rivalry, perceptual dominance is determined primarily by low-level features such as the size or relative width of the face, or salient local features such as birthmarks.

Lisa Kossmann

Talks and Poster Presentations