Abstract
We often consider a visual space to be cluttered if it contains a large number of items that are crowded and untidy. However, measuring clutter is complex, and clutter has been quantified using objective measures, such as mathematical models, most of which capture different aspects of clutter, and subjective measures, such as human judgments. In this study, using a well-controlled semi-realistic image database, we examined the contributions of visual factors, such as the number of items in a scene, their spatial organization, and the complexity of the surface where items are placed, to clutter by testing the predictive performance of clutter metrics and collecting human judgments of perceived clutter with respect to these factors. Our results revealed that, although all factors contribute to clutter in visual scenes, estimates by humans and metrics account only for a subset of these factors, suggesting that visual clutter is a multifaceted construct, and subjective and objective clutter should be considered together to account for several of its dimensions as a more integrative measure.
Keywords: visual clutter, subjective clutter, scene perception
Introduction
Everyday, we encounter visual clutter through cluttered rooms, desks, websites, and so on. Clutter can affect the way we perceive our visual environment. Even a simple task, such as looking for a pen on a desk, can be time consuming if the desk is disorganized and cluttered with many objects. In contrast, on an organized desk, finding the pen could be trivial. Just as clutter makes simple visual tasks more difficult, clutter and visual complexity can also negatively impact performance in more demanding settings, such as radiology (Drew et al., 2013), baggage screening (Biggs & Mitroff, 2015), or driving (Horrey & Wickens, 2004), which often require decision-making in complex visual environments. For example, during the diagnostic process, radiologists routinely interpret complex medical images that often contain dense tissues, anatomical structures, or imaging artifacts, which can interfere with the detection of abnormalities, such as nodules or masses, by introducing clutter.
Clutter is relevant in various fields, such as psychology, engineering, human–computer interaction, and marketing, where it significantly affects human performance by increasing information load and often making the identification of relevant information more difficult (Rosenholtz, Li, Mansfield, & Jin, 2005; Rosenholtz, Li, & Nakano, 2007). Designers use strategies to reduce clutter for easy processing of visual information (Tufte, 2001). Developing a deeper understanding of visual clutter can enhance our knowledge of the mechanisms of the visual system and inform design decisions. But what exactly is clutter?
Visual clutter is an intuitive concept. We often consider a space cluttered if it has a lot of items that are crowded and untidy. Despite its intuitive nature, quantifying clutter in a visual scene can be challenging (Rosenholtz et al., 2007; Semizer & Michel, 2019). For example, assume that you are given a desk surface with multiple objects scattered across it and are asked to estimate the amount of clutter in this space. How would you go about this task? Would you just count the number of objects? If so, how would you define an object? Is a book an object? What if the book has text or pictures on the cover? Would you count each word, letter, or stroke as an object? What if there were texture elements on the desk surface? Should you consider them in your estimate? Quantifying clutter is more complex than what our intuition suggests. Despite this challenge, a considerable amount of research has sought to define and measure clutter in visual scenes.
Visual clutter is associated with several factors, including the number of items (Palmer, 1994; Palmer, 1995), the density of times (Mack & Oliva, 2004; van den Berg, Cornelissen, & Roerdink, 2009), the layout or organization of the items (Bravo & Farid, 2008; Rosenholtz et al., 2005; van den Berg et al., 2009), and the similarity of the items to a target object or category in a scene (Semizer & Michel, 2022). Such clutter definitions focus mainly on scene properties. Alternatively, clutter can be defined in terms of task performance, in addition to the structure and density of the items. Based on such a framework, visual clutter refers to “the state in which excess items, or their representation or organization, lead to a degradation of performance at some task” (Rosenholtz et al., 2005, p. 761). Thus, the definition of clutter significantly impacts how it is quantified. To date, various methods have been proposed to quantify clutter, ranging from objective methods, such as mathematical models, to subjective methods, such as human judgments of clutter (Moacdieh & Sarter, 2015).
Objective clutter metrics measure clutter in visual scenes using a wide range of methods, each capturing different aspects of visual complexity. We first note that recent advances in artificial intelligence and computer vision have greatly increased our understanding of visual features (Fan, Zhang, Xu, Cheng, & Shao, 2022). However, this work is most concerned with clutter metrics within the psychophysics literature. Feature-based clutter metrics, such as edge density (Mack & Oliva, 2004), feature congestion (Rosenholtz et al., 2005; Rosenholtz et al., 2007), and subband entropy (Rosenholtz et al., 2007), quantify clutter based on the density and variability of visual features. This approach relates to summary statistics accounts of peripheral vision (Rosenholtz, 2016). Spatial organization metrics, such as segmentation-based clutter metrics (Bravo & Farid, 2008) and proto-object models (Yu, Samaras, & Zelinsky, 2014), quantify clutter based on the spatial arrangement and global structure of the scene. In addition, target-specific metrics, such as those proposed by Semizer and Michel (2022), quantify clutter in relation to a specific object, focusing on task-relevant effects of clutter. Together, these metrics have previously been shown to predict performance in visual search tasks, which involve searching for target objects in visual scenes (Bravo & Farid, 2008; Henderson, Chanceaux, & Smith, 2009; Mack & Oliva, 2004; Neider & Zelinsky, 2011; Rosenholtz et al., 2005; Rosenholtz et al., 2007; Semizer & Michel, 2019; Semizer & Rosenholtz, 2025), suggesting that performance declines with increasing clutter.
Clutter can also be measured using subjective ratings from humans, mainly to validate estimates given by objective clutter metrics. Studies using a wide range of natural or synthetic images with varying complexity showed that human clutter judgments were correlated with model estimates (e.g., Bravo & Farid, 2008; Henderson et al., 2009; Neider & Zelinsky, 2011; Yu et al., 2014). In these studies, participants were given a set of images and asked to provide judgments on perceived clutter using methods such as numerical ratings, rankings, and categorization (e.g., high/low clutter). Together, these findings demonstrate that human clutter ratings can be captured using computational clutter metrics, suggesting evidence for the validity of these metrics in quantifying perceived clutter.
Although the aforementioned metrics clearly capture different aspects of clutter, there has been little investigation into the systematic manipulation of the image characteristics that lead to it. The purpose of the current study was to investigate the factors that contribute to perception of visual clutter and to evaluate the performance of clutter metrics with respect to these factors. To this end, we generated a well-controlled semi-realistic dataset of tabletop images featuring various object arrangements. We systematically varied three main potential factors, some of which have previously been linked to clutter: item quantity, spatial organization, and surface complexity. For item quantity, we manipulated the number of items placed on table surfaces. For spatial organization, we manipulated the arrangement of items on the tables. For surface complexity, we manipulated the texture elements of the table surfaces. We collected human ratings to assess perceived clutter and applied clutter metrics to these images to quantify objective clutter. We tested which factors effectively predict the judgments of human participants and the performance of clutter metrics. Our results revealed that, although all factors contribute to clutter in visual scenes, estimates by humans and clutter metrics account only for a subset of these factors, suggesting that clutter is a multifaceted construct, with different aspects captured by a range of measures.
Methods
Image database creation
Our database contains 400 images depicting virtual scenes of tables with objects that vary in item quantity, spatial organization, and surface complexity (Figure 1). We used tabletops as our primary paradigm, because they represent a common real-world context in which cluttered visual scenes occur naturally. Indeed, tabletop images have previously been used to study the role of background complexity in visual search tasks (Wolfe, Oliva, Horowitz, Butcher, & Bompas, 2002). Although our study does not focus on visual search, this paradigm offers an ecologically valid context for studying the perception of clutter.
Figure 1.
Example images from the image database with varying spatial organization, surface complexity, and item quantity. For surface complexity, images were selected to represent the low and high ends of the distribution).
Images were created using Blender (Community, 2018), a virtual three-dimensional (3D) modeling software that provides a variety of tools for 3D modeling, texturing, and lighting, allowing users to create everything from simple geometric shapes to intricate items such as books, plates, and electronic gadgets. Blender also supports the import of premade 3D models, which can be placed manually or programmatically to create a cluttered table scene.
Images of the tables were captured from a consistent camera position and frame, with static lighting conditions, so that all of the table and objects were visible in the rendered image. In terms of resolution and other features, all images were captured with the same settings. The objects placed on the tables were selected to vary in type, size, and shape. Examples objects include laptops, pens, table lamps, books, apples, and so on. Objects were placed in various permutations, from being positioned on the table, vertically stacked, and so on. We specifically manipulated three factors: item quantity, spatial organization, and surface complexity.
Item quantity
We used 20 different sets of items, with one-half containing approximately 30 objects (high) and the other one-half containing approximately 10 objects (low). Each set was placed on all possible table surfaces.
Spatial organization
Spatial organization was manipulated through spatial alignment of the objects on the tables. In the organized condition, objects were placed parallel to the table, resulting in a structured layout with minimal occlusions between objects. In the disorganized condition, objects were placed at various angles relative to each other and the table, resulting in a less structured layout, which could increase the likelihood of occlusions between objects. For each organized table with a given table surface and set of items, a disorganized version was created. Categories were validated in a separate study (see Perceived spatial organization).
Surface complexity
We varied the texture and material appearance of the table surfaces. Examples include wooden or marble tables in a range of colors, such as red, green, or gray. We used 10 different textures and quantified the complexity of table surfaces using human judgments (see Perceived surface complexity).
Stimulus validation and characterization
To validate and characterize the properties of stimuli in our image database, we ran two supplementary experiments the first to verify the categorization of spatial organization, and the second to characterize the complexity of table surfaces. We describe each experiment in detail.
Perceived spatial organization
Participants
A total of 60 participants (25 female, 35 male; M age = 28.6 years, SD = 4.6) with normal or corrected-to-normal vision, none of whom participated in any other experiment, were recruited online from Prolific (www.prolific.co). Data from one participant were not saved due to a technical error, leaving 59 participants for the analysis. All participants provided informed consent and received compensation for their participation in the study.
Stimuli and apparatus
The stimuli consisted of 40 images, representing organized and disorganized arrangements of each set of objects, presented on a table surface with medium complexity, as determined in a separate experiment (see Perceived surface complexity for details). Each participant rated only one-half of the stimuli, with each set presented only once per participant. Across participants, each image was rated by one-half of the sample. The task was programmed in PsychoPy/PsychoJS (Peirce, Gray, Simpson et al., 2019) and completed online, with participants using their own desktop computers.
Procedure
At the start of the task, participants were asked to rate the extent to which each image appeared disorganized by clicking on a 10-point scale ranging from “very organized” to “very disorganized.” Participants completed 4 practice trials to familiarize themselves with the procedure, followed by 20 experimental trials. The task was self-paced and took less than 10 minutes to complete.
Perceived surface complexity
Participants
A new group of 60 participants (29 female, 31 male; M age = 28.8 years, SD = 4.1) with normal or corrected-to-normal vision, independent of those in the previous experiment, was recruited online from Prolific (www.prolific.co). Participants received compensation for their participation in the study. Ten participants were excluded from the final sample, resulting in a total of 50 participants, based on the following criteria. Participants were asked to report their classification and decision-making method at the end of the experiment. Seven participants reported ranking the images according to their visual appeal and their own liking, one participant reported randomly ranking the images, and one participant did not respond to the question. All participants provided their informed consent before participating in the study.
Stimuli and apparatus
The stimuli consisted 10 images of empty tables with varying levels of surface complexity that were used in the image database. The task was programmed in PsychoPy/PsychoJS (Peirce et al., 2019). Participants completed the task on their own desktop computers.
Procedure
At the start of the experiment, participants were given instructions to classify the images according to their perceived surface complexity by dragging and dropping each image into an empty box numbered from 1 to 10. Larger numbers indicated higher complexity (Figure 2a). After participants completed the classification, they logged their responses by clicking a button. Participants were then asked to report their classification and decision-making method by entering text in a text box.
Figure 2.
(A) Response screen for the ranking task on the left. Participants were asked to rank the images from 1 to 10 (low to high) based on perceived surface complexity by dragging and dropping the images into empty boxes. (B) Response screen for the rating task. Participants were asked to rate the perceived clutter of the images by clicking on a 10-point scale.
To familiarize themselves with the drag-and-drop response, participants completed a practice trial in which they were asked to drag and drop an image into a box. On average, the experiment took less than 5 minutes to complete.
Human clutter estimation
Participants
Power analysis indicated a minimum of 70 participants were required to detect all main effects with a moderate effect size (Cohen’s f = 0.25) at 90% power.1 To account for possible exclusions, a total of 100 participants (65 female, 34 male, 1 undisclosed; M age = 22.5 years, SD = 2.2) with normal or corrected-to-normal vision, none of whom participated in any other experiment, were recruited online from Prolific (www.prolific.co). Participants received compensation for their participation in the study. Data from two participants were not saved due to a technical error. Twelve participants did not fully complete an initial practice task. Fourteen participants did not meet the performance criteria in catch trials (see Procedure for details). Data from these participants were excluded from the analysis, resulting in a total of 72 participants in the final sample. All participants provided their informed consent before participating in the study.
Stimuli and apparatus
The stimuli consisted of 400 images from our image dataset. In addition, empty tables with six different texture elements, each presented twice, were used on the catch trails to check for data quality. The surface texture of the catch images was different from the texture in the experimental images. The task was programmed in PsychoPy/PsychoJS (Peirce et al., 2019). Participants completed the task on their own desktop computers.
Procedure
At the start of the experiment, participants were given instructions about the task and asked to judge how much visual clutter each image contained. We intentionally did not provide a definition of “clutter” to avoid biasing responses and to allow participants to apply their own interpretation. During this initial phase, to familiarize participants with the extent of the amount of clutter in our stimulus set, we presented nine images, eight representing different conditions, and one showing an empty table, and asked participants to rank the images based on visual clutter as follows: “Please rank these images from least cluttered to most cluttered. The image you consider most cluttered should receive the highest rank.” The images used in this stage were not part of the database, but resembled those in the database.
After completing the initial phase, participants moved onto the main task, where they were instructed as follows: “For each image, please rate the amount of visual clutter by clicking on the slider at your desired tick mark, ranging from ‘low clutter’ to ‘high clutter.”’ Responses were recorded using a slider on a 10-point scale (Figure 2b). There were 400 experimental trials. To verify data quality, we also included 12 catch trials in which an empty table was presented. The placement of the catch trials was semi-random, occurring approximately every 33 trials with some jitter (±2–4). Participants who rated more than two catch trials with a score of greater than 3 were excluded from the analysis. After every 100 trials, participants were given the choice to take a break. The task was self-paced and on average took approximately 30 minutes.
Model-based clutter estimation
To estimate the amount of clutter in the images in our database, we applied four computational clutter metrics: feature congestion (Rosenholtz et al., 2005; Rosenholtz et al., 2007), subband entropy (Rosenholtz et al., 2007), edge density (Mack & Oliva, 2004), and a segmentation-based metric (Bravo & Farid, 2008). We chose these metrics because of their widespread use in the field and the availability of their code. These metrics have been successful in predicting human performance, mainly in visual search tasks, as a function of clutter (Bravo & Farid, 2004; Bravo & Farid, 2008; Henderson et al., 2009; Neider & Zelinsky, 2011; Rosenholtz et al., 2007; Semizer & Michel, 2019). We provide brief description of each metric.
The feature congestion metric (Rosenholtz et al., 2005; Rosenholtz et al., 2007) quantifies clutter as a function of local variance in visual features such as orientation, color, and contrast, with the underlying assumption that displays with greater local variability correspond with greater clutter. The subband entropy metric (Rosenholtz et al., 2007) computes clutter by estimating the number of bits required to encode an image while using steerable pyramids (Simoncelli & Freeman, 1995) to decompose images, assuming that more cluttered displays require more bits to encode. The edge density metric (Mack & Oliva, 2004) estimates the amount of clutter by computing the density of high-contrast edges in each image, assuming that more cluttered displays have more edges. The segmentation-based clutter metric (Bravo & Farid, 2008) measures clutter using an image segmentation algorithm (Felzenszwalb & Huttenlocher, 2004), based on the assumption that the more cluttered an image is, the more difficult it is to segment to distinct regions. We used a modified version of this metric in which clutter was computed as the geometric mean of multiple estimates obtained by sampling various regions within each image (Semizer & Michel, 2019).
Data analysis
Clutter models were implemented in MATLAB software (MathWorks, Natick, MA). All statistical analyses were conducted in R (R Core Team, 2024). GLMs were fitted using the base R glm() function and linear mixed-effects models were fitted using the lme4 R package (Bates, Mächler, Bolker, & Walker, 2015).
Results
We discuss our results as follows. First, we present the results of the perceived spatial organization experiment in which participants were asked to judge the extent to which the item arrangements appeared disorganized. Next, we report the results of the perceived complexity experiment in which participants were asked to judge the complexity of table surfaces. Then, we discuss the results of the perceived clutter experiment in which participants were asked to judge the clutter in the images, followed by the performance of computational clutter metrics in accounting for three factors manipulated in our image set. Finally, we report results examining the relationship between human judgments and computational metrics of clutter.
Perceived spatial organization
Figure 3 shows the average disorganization ratings in each level of spatial organization and item quantity. The agreement among participants’ ratings was evaluated using an intraclass correlation coefficient estimated from an intercept-only linear mixed-effects model with random intercepts for participants and images. The intraclass correlation coefficient was 0.72, suggesting a consistent agreement among the ratings of the participants.
Figure 3.

Spatial organization ratings as a function of item quantity conditions. Gray points (slightly jittered) indicate individual image responses, and the lines represent average ratings in the low and high item quantity conditions. Shaded areas represent 95% confidence intervals.
Ratings were analyzed using a linear mixed-effects model including spatial organization, item quantity, and their interactions. Spatial organization and item quantity were entered as fixed effects, with random intercepts and slopes for both predictors by participant, and random intercepts by image. The model revealed significant effects of spatial organization (β = 3.927; SE = 0.305; t = 12.874; p < 0.001), item quantity (β = 0.716; SE = 0.284; t = 2.523; p = 0.015), and their interaction (β = 1.955; SE = 0.373; t = 5.245; p < 0.001). Images labeled as disorganized received higher disorganization ratings than those labeled as organized, and this effect was larger in sets with more items. Based on this validation, we used spatial organization as a categorical factor (organized vs. disorganized) in our main experiment, because it directly represents the theoretical categorization imposed by our manipulation.
Perceived surface complexity
Figure 4 shows the order of images based on perceived surface complexity. A Kendall’s coefficient of concordance was calculated to evaluate the inter-rater agreement, revealing a statistically significant level of agreement among the participants, W = 0.48, χ2(9) = 214.34, p < 0.001. Participants reported that they based their decisions on the pattern, texture, and color of the surface of the table.
Figure 4.
Empty tables in raster scan order based on their perceived surface complexity rank.
Perceived clutter
Figure 5 shows the average clutter values by human participants in each condition. To examine the consistency among participants in their clutter judgments, a two-way random-effects model, measuring absolute agreement among participants, was implemented. The intraclass correlation coefficient for the average ratings was 0.98, F(399, 1648) = 70.80, p < 0.001, suggesting a high level of consistency between the ratings.
Figure 5.

Average clutter values from human participants as a function of surface complexity, item quantity, and item organization. Lines represent model predictions. Shaded areas indicate 95% intervals. Points represent average participant responses.
To understand the effects of item quantity, spatial organization, and surface complexity on clutter values, a linear mixed-effects model was implemented. The model included the main effects of the three factors and their interactions. Item quantity, spatial organization, and surface complexity were entered as fixed effects. Random intercepts were included for participants and images, and by-participant random slopes were specified for item quantity and spatial organization. Estimated marginal means, estimated means adjusted for other factors, were also computed to aid interpretation.
As shown in Table 1, the results indicated a main effect of item quantity, β = −2.683, SE = 0.261, t(383.03) = −10.295, p < 0.001, suggesting that images with more items (M = 7.05, SE = 0.10) were rated as more cluttered than those with fewer items (M = 3.98, SE = 0.07). There was also a main effect of spatial organization, β = 1.941, SE = 0.264, t(369.58) = 7.355, p < 0.001, indicating that disorganized images (M = 4.87, SE = 0.11) were rated as more cluttered than organized images (M = 6.17, SE = 0.13). A significant item quantity and spatial organization interaction was also found, β = −1.021, SE = 0.306, t(391.99) = −3.337, p < 0.001, indicating a larger effect of spatial organization in images containing more items than in those containing fewer items. None of the other effects reached significance (p > 0.05).
Table 1.
Linear mixed-effects model coefficients predicting human clutter ratings.
| β (SE) | t | p | |
|---|---|---|---|
| Quantity | −2.683 (0.261) | −10.295 | <0.001 |
| Complexity | 0.041 0.026) | 1.567 | 0.118 |
| Organization | 1.941 (0.264) | 7.355 | <0.001 |
| Quantity × Complexity | 0.008 (0.037) | 0.212 | 0.832 |
| Quantity × Organization | −1.021 (0.306) | −3.337 | <0.001 |
| Complexity × Organization | −0.040 (0.037) | −1.079 | 0.281 |
| Quantity × Complexity × Organization | 0.03 (0.0524) | 0.622 | 0.534 |
Note: Bolded coefficients indicate p < 0.05.
Clutter metric predictions
To compare how clutter metrics perform relative to each other, we ran a series of correlational analyses across all images using Spearman’s method. As shown in Table 2, several significant positive correlations were observed among clutter metrics. Feature congestion was strongly correlated with both edge density (ρ = 0.77; p < 0.001) and segmentation (ρ = 0.70; p < 0.001) whereas it was weakly correlated with subband entropy (ρ = 0.20, p < 0.001). In addition, edge density was strongly correlated with segmentation (ρ = 0.77; p < 0.001). subband entropy had a very weak correlation with edge density (ρ = 0.11; p < 0.05).
Table 2.
Spearman correlation coefficients among feature congestion (Rosenholtz et al., 2005; Rosenholtz et al., 2007), subband entropy (Rosenholtz et al., 2005), edge density (Mack & Oliva, 2004), and segmentation-based (Bravo & Farid, 2008) clutter metrics.
| Feature congestion | subband entropy | Edge density | Segmentation | |
|---|---|---|---|---|
| Feature congestion | 1.00 | – | – | – |
| subband entropy | 0.20*** | 1.00 | – | – |
| Edge density | 0.77*** | 0.11* | 1.00 | – |
| Segmentation | 0.70*** | 0.06 | 0.77*** | 1.00 |
Note: *p < 0.05, **p < 0.01, ***p < 0.001.
Next, to assess the effects of item quantity, spatial organization, and surface complexity on clutter measures, we ran a series of generalized linear models (GLMs) with Gaussian distribution on each of the clutter metrics, including all main effects and their interactions. In cases of non-normal data, transformations were applied. Estimated marginal means were computed to facilitate interpretation. Tables 3 and 4 show coefficients for GLMs with three variables as predictors, as well as their interactions. Figure 67 shows predicted clutter values for each metric based on GLMs, with the effects of three variables. Model predictions are back-transformed where necessary and presented on the original scale.
Table 3.
GLM coefficients predicting clutter values across feature congestion and subband entropy metrics.
| Feature congestion | subband entropy | |||||
|---|---|---|---|---|---|---|
| Predictor | β (SE) | t | p | β (SE) | t | p |
| Quantity | −0.102 (0.021) | −4.815 | <0.001 | 0.031 (0.051) | 0.611 | 0.542 |
| Complexity | 0.006 (0.003) | 2.317 | 0.021 | −0.012 (0.006) | −1.942 | 0.053 |
| Organization | −0.001 (0.021) | −0.038 | 0.97 | 0.004 (0.051) | 0.071 | 0.944 |
| Quantity × Complexity | 0.003 (0.004) | 0.69 | 0.492 | −0.001 (0.009) | −0.217 | 0.829 |
| Quantity × Organization | −0.008 (0.030) | −0.262 | 0.793 | 0.007 (0.071) | 0.093 | 0.9260 |
| Complexity × Organization | −0.007 (0.004) | −0.153 | 0.878 | 0.003 (0.009) | 0.386 | 0.699 |
| Quantity × Complexity × Organization | 0.001 (0.005) | 0.090 | 0.928 | −0.003 (0.012) | −0.208 | 0.835 |
| Akaike information criterion | −1240.8 | −544.7 | ||||
| 0.465 | 0.031 | |||||
Note: Bolded coefficients indicate p < 0.05.
Table 4.
GLM coefficients predicting clutter values across edge density and segmentation metrics.
| Edge density | Segmentation | |||||
|---|---|---|---|---|---|---|
| Predictor | β (SE) | t | p | β (SE) | t | p |
| Quantity | −0.667 (0.131) | −5.102 | <0.001 | −33339.6 (9173.6) | −3.634 | <0.001 |
| Complexity | 0.065 (0.016) | 4.138 | <0.001 | 5661.6 (1109.5) | 5.103 | <0.001 |
| Organization | 0.125 (0.131) | 0.953 | 0.341 | −8007.0 (9173.6) | −0.873 | 0.383 |
| Quantity × Complexity | 0.060 (0.022) | 2.687 | 0.008 | 2212.4 (1569.1) | 1.410 | 0.159 |
| Quantity × Organization | −0.167 (0.185) | −0.902 | 0.367 | −11619.3 (12973.5) | −0.896 | 0.371 |
| Complexity × Organization | −0.007 (0.022) | −0.295 | 0.768 | −370.4 (1569.1) | −0.236 | 0.814 |
| Quantity × Complexity × Organization | 0.007 (0.032) | 0.235 | 0.815 | 559.5 (2219.1) | 0.252 | 0.801 |
| Akaike information criterion | 216.61 | 9143.9 | ||||
| 0.443 | 0.449 | |||||
Note: Bolded coefficients indicate p < 0.05.
Figure 6.
Predicted clutter values for each metric across as a function of surface complexity, item quantity, and item organization for four metrics. (A–D) Feature congestion, subband entropy, edge density, and segmentation metrics, respectively. Lines represent model estimates. Shaded bands indicate 95% confidence intervals. Note that the y-axes are offset.
Figure 7.
Scatter plots showing the relationship between clutter values predicted by the computational metrics and human judgments. (A–D) Feature congestion, subband entropy, edge density, and segmentation metrics, respectively. Each gray point represents an image, and the trend lines show the linear fits. Correlations were assessed using Spearman’s method. Note that the y-axes are offset.
For the feature congestion metric, a log transformation was performed due to positive skewness of the data. The results showed a main effect of surface complexity, β = 0.006, SE = 0.003, t(392) = 2.317, p < 0.05, indicating a small but significant increase in clutter values as complexity increases. There was a main effect of item quantity, β = −0.102, SE = 0.021, t(392) = −4.815, p < 0.001. Clutter values were higher in images with more items (M = 2.05, SE = 0.01) than those with fewer items (M = 1.88, SE = 0.01). The other main effect and all interaction effects were not significant (all ps > 0.05).
For the subband entropy metric, the distribution of clutter values were negatively skewed, so a reflected log transformation was performed. None of the effects were significant, although the main effect of surface complexity approached significance, β = −0.012, SE = 0.006, t(392) = −1.942, p = 0.053.
For the edge density metric, a log transformation was performed because of the positive skewness of the data. The results revealed a main effect of surface complexity, β = 0.065, SE = 0.0016, t(392) = 4.138, p < 0.001, indicating that, as complexity increased, clutter also increased. There was a main effect of item quantity, β = −0.667, SE = 0.131, t(392) = −5.102, p = 0.001. Images with more items (M = 0.013, SE = 0.001) had higher clutter than those with fewer items (M = 0.009, SE = 0.001). The interaction effect between item quantity and surface complexity was also significant, B = 0.060, SE = 0.022, t(392) = 2.687, p = 0.008. The effect of surface complexity was greater for images with fewer items compared with those with more items. None of the other effects were significant (all ps > 0.05).
For the segmentation-based metric, there was a main effect of surface complexity (β = 5661.6; SE = 1109.5; t(392) = 5.103; p < 0.001), suggesting that higher complexity leads to increased clutter. The main effect of item quantity was also significant, β = −33339.6, SE = 9173.6, t(392) = −3.634, p < 0.001. Higher clutter values were observed in images with more items (M = 149, 855, SE = 1, 560) compared with those with fewer items (M = 124, 412, SE = 1, 560). None of the remaining effects reached significance (all ps > 0.05).
Perceived clutter versus clutter metric predictions
To examine the degree to which clutter metrics correspond with perceived clutter as rated by humans, we ran a series of correlational analysis using Spearman’s method. Human ratings were significantly correlated with the feature congestion metric (ρ = 0.60; p < 0.001), the edge density metric (ρ = 0.60; p < 0.001), and the segmentation-based metric (ρ = 0.40; p < 0.001), but not with the subband entropy metric (p > 0.05).
To further examine how metrics predict human clutter judgments, we implemented a linear regression model with four metrics as predictors. The full model explained 45% of the variance in human judgments, F(4, 395) = 80.07, p < 0.001. Specifically, the feature congestion metric was the strongest predictor (β = 8.508; SE = 0.613; t = 13.891; p < 0.001), explaining 39% of the variance (sr2 = 0.39). The edge density (β = 57.81; SE = 17.56; t = 3.293; p < 0.01), and segmentation (β = −1.61 × 10−5; SE = 3.41 × 10−6; t = −4.707; p < 0.001), metrics contributed to 16.9% and 13.1% of the variance, respectively. Although the subband entropy metric was a significant predictor (β = −1.814; SE = 0.340; t = −5.345; p < 0.001), its contribution was negligible (rs2 < 0.001).
Discussion
Intuitively, when a visual space has a lot of “stuff” in a disorderly fashion, we talk about clutter or, better yet, decluttering the space. But which specific visual factors cause visual clutter? The purpose of the current study was to investigate the visual factors that contribute to clutter. In particular, we measured human perception of clutter and tested the performance of computational clutter metrics across a set of well-controlled, semi-realistic images manipulated by three visual factors: item quantity, spatial organization, and surface complexity. Although each clutter metric captures an important aspect of visual complexity, our analyses revealed both consistencies and discrepancies in their predictions relative to human judgments. Notably, when all metrics and human data are considered together, clutter increases with surface complexity, even when object count is unchanged. Surface complexity has a stronger effect with fewer items, and disorganization becomes more important when many objects are present.
A cluttered space is often expected to contain a large number of objects. Given two visual scenes with similar features, one containing a higher number of objects would be considered to be more cluttered. In line with this expectation, our results showed that participants judged tables that contain more items to be more cluttered. Similarly, the number of items placed on a table was a significant predictor of the performance of three metrics, namely, the feature congestion metric (Rosenholtz et al., 2005; Rosenholtz et al., 2007), the edge density metric (Mack & Oliva, 2004), and the segmentation-based metric (Bravo & Farid, 2008). The clutter values estimated by these metrics increased when the tables contained a large number of items. As the number of items in a scene increases, the density of features such as edges and textures tends to increase, which likely contributes to the observed effects. The lack of effect in the subband entropy measure might be due to its sensitivity to the distribution of information, its complexity and variability, across a scene structure at different scales, rather than the count of items in the scene. Together, these results suggest that the number of items in a scene is a strong predictor of objective and subjective measures of visual clutter. This finding provides further evidence for the view that clutter is associated with object density (e.g., Mack & Oliva, 2004; van den Berg et al., 2009) and is especially important in the context of visual search where set size (Palmer, 1994; Palmer, 1995) is a significant contributor of performance (Wolfe, 1998).
Decluttering a desk often involves placing objects that are similar to each other in close proximity (e.g., putting books and notepads in a pile) and organizing them spatially. If a scene is highly disorganized, we would consider it to be more cluttered. In this regard, the arrangement of objects in a scene can contribute to clutter (Bravo & Farid, 2004; Rosenholtz et al., 2005; van den Berg et al., 2009). In fact, our results from the perceived clutter experiment revealed that participants rated images as more cluttered when items were disorganized. However, spatial organization did not significantly predict clutter estimates given by the metrics. The dissociation between clutter estimates by metrics and perceived clutter by humans might be due to the nature of our manipulation to achieve disorganization, which was done by altering the arrangement of the objects relative to each other and the table, but did not necessarily involve grouping. It is possible that our manipulation, which focuses on relative positioning, was at a higher level than what these metrics focus on.
Clutter can also be the result of the complexity of visual elements in a scene. If a table has the same number of objects arranged in the same manner, we expect to see more clutter when the table has a patterned surface rather than a plain surface. In line with this expectation, surface complexity, defined as the perceived complexity of the surface of the table as judged by human participants, predicted the performance of most clutter metrics. The clutter estimates were higher when the surface of the table was perceived to be more complex, suggesting that surface complexity contributes significantly to objective clutter measures. The lack of a complexity effect in human clutter judgments suggests that complexity does not reliably influence perceived clutter under the present conditions, although the measurement approach may reduce sensitivity to subtle perceptual differences.
Our findings also revealed an interaction between item quantity and surface complexity, based on the edge–density clutter metric (Mack & Oliva, 2004), suggesting larger effects of surface complexity when images contained fewer items. This is likely due to the fact that, when there are few items on a table, surface elements become more visible, and thus the complexity effects become more pronounced.
Our analysis of perceived clutter judgments by human participants revealed additional interaction effects between these factors. The observed spatial organization effect, where disorganized images were rated as more cluttered than organized ones, was more pronounced at higher item quantities. This finding indicates that, when tables are crowded with many items, the effect of organization is amplified; however, when items are sparse, the effect is still there, but reduced. This is likely due to our manipulation of spatial organization. The disorganization in low item quantity was achieved mostly by changing the arrangement of items on the tables to make sure that a similar amount of surface was covered between the organized and disorganized conditions. We suspect that this also contributes to the lack of spatial organization effects in most metrics.
Some clutter metrics were strongly related to each other, whereas others showed weak or no relationship. Specifically, the feature congestion metric (Rosenholtz et al., 2005; Rosenholtz et al., 2007) and the edge density metric (Mack & Oliva, 2004) were strongly related to the segmentation-based metric (Bravo & Farid, 2008); however, the subband entropy metric (Rosenholtz et al., 2007) seemed to have a less consistent relationship with the other metrics. The performance of these metrics has been compared in previous studies. For example, Semizer and Michel (2019) applied these metrics to a set of natural images and reported strong correlations between all the metrics. In addition to the differences between the nature of the stimuli between the current and the previous study, the distribution of clutter values in the current study were not normal and highly skewed in some cases, which could explain the differences observed between the two studies. Furthermore, our findings showed a strong to moderate relationship between the predictions of the clutter metrics and the perceived clutter judged by humans, suggesting that these clutter metrics were effective in representing perceived clutter to some degree. Indeed, clutter metrics accounted for a significant portion of the variance in perceived clutter judgments. Previously, using rural, urban, and suburban scenes created with a video game, Neider and Zelinsky (2011) reported that objective clutter, measured by the edge density metric, correlated with subjective clutter, measured by collecting human ratings.
Although these findings contribute to our understanding of what causes visual clutter, our study is limited in some aspects. First, we have a very specific and controlled image database composed of objects placed on tabletops viewed from a particular angle. The objects are simple items, and the visual elements, such as color, contrast, or texture, do not have a high degree of variability. The manipulation of spatial organization only involved changing the arrangement of objects and was agnostic to perceptual grouping. Although we purposely designed our images in this manner to control for specific factors, more complex images should also be tested to further examine the dimensions that influence clutter in scenes. Second, although our image database is composed of images that appear realistic, they are created by us and, therefore, highly controlled. Examining how the factors tested here contribute to clutter estimates in natural images would provide further evidence to support their role in visual clutter. Moreover, although we controlled the image characteristics relevant to our study, we did not control the image size on participants’ computers and let it vary naturally. We chose this approach to prioritize external validity; however, future studies could control image size to examine its effects on perceived clutter. Third, our work does not incorporate visual features inferred from neural networks, which have been shown to be useful in computer vision for tasks such as object detection, which may relate to clutter measurement (Fan et al., 2022). We leave such investigation for future work. Last, we only used a subset of existing clutter metrics; future work could incorporate more recent metrics.
Conclusions
Our study suggests that, although the number of items in a scene, their spatial organization, and the complexity of the surface where items are placed contribute to clutter in visual scenes, human judgments and existing clutter metrics account only for a subset of these factors. This finding suggests that visual clutter is a multifaceted construct, and subjective and objective clutter should be considered together to account for several of its dimensions as a more integrative measure. Future studies examining additional visual elements can help us unpack what constitutes clutter and inform more effective design choices.
Acknowledgments
Commercial relationships: none.
Corresponding author: Yelda Semizer.
Email: yelda.semizer@njit.edu.
Address: Department of Humanities and Social Sciences, New Jersey Institute of Technology, Newark, NJ 07102, USA.
Footnotes
One of the variables, surface complexity, was later treated as continuous; therefore, the power achieved for this variable may differ from the initial simulation.
References
- Bates, D., Mächler, M., Bolker, B., & Walker, S. (2015). Fitting linear mixed-effects models using lme4. Journal of Statistical Software, 67(1), 1–48, 10.18637/jss.v067.i01. [DOI] [Google Scholar]
- Biggs, A. T., & Mitroff, S. R. (2015). Improving the efficacy of security screening tasks: A review of visual search challenges and ways to mitigate their adverse effects. Applied Cognitive Psychology, 29(1), 142–148, 10.1002/acp.3083. [DOI] [Google Scholar]
- Bravo, M. J., & Farid, H. (2004). Search for a category target in clutter. Perception, 33(6), 643–652, 10.1068/p5244. [DOI] [PubMed] [Google Scholar]
- Bravo, M. J., & Farid, H. (2008). A scale invariant measure of clutter. Journal of Vision, 8(2008), 23.1–9, 10.1167/8.1.23. [DOI] [PubMed] [Google Scholar]
- Community, B. O. (2018). Blender - a 3D modelling and rendering package [Computer software manual]. Amsterdam: Stichting Blender Foundation. Retrieved from http://www.blender.org. [Google Scholar]
- Drew, T., Vo, M. L.-H., Olwal, A., Jacobson, F., Seltzer, S. E., & Wolfe, J. M. (2013). Scanners and drillers: Characterizing expert visual search through volumetric images. Journal of Vision, 13(10), 3. Retrieved from 10.1167/13.10.3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Fan, D.-P., Zhang, J., Xu, G., Cheng, M.-M., & Shao, L. (2022). Salient objects in clutter. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(2), 2344–2366. [DOI] [PubMed] [Google Scholar]
- Felzenszwalb, P. F., & Huttenlocher, D. P. (2004). Efficient graph-based image segmentation. International Journal of Computer Vision, 59(2), 167–181, 10.1023/B:VISI.0000022288.19776.77. [DOI] [Google Scholar]
- Henderson, J. M., Chanceaux, M., & Smith, T. J. (2009). The influence of clutter on real-world scene search: Evidence from search efficiency and eye movements. Journal of Vision, 9(1), 1–8, 10.1167/9.1.32. [DOI] [PubMed] [Google Scholar]
- Horrey, W. J., & Wickens, C. D. (2004). Driving and side task performance: The effects of display clutter, separation, and modality. Human Factors, 46(4), 611–624. [DOI] [PubMed] [Google Scholar]
- Mack, M. L., & Oliva, A. (2004). Computational estimation of visual complexity. Poster presented at the 12th Annual Object, Perception, Attention, and Memory Conference, Minneapolis, MN. [Google Scholar]
- Moacdieh, N., & Sarter, N. (2015). Display clutter: A review of definitions and measurement techniques. Human Factors, 57(1), 61–100, 10.1177/0018720814549312. [DOI] [PubMed] [Google Scholar]
- Neider, M. B., & Zelinsky, G. J. (2011). Cutting through the clutter: Searching for targets in evolving complex scenes. Journal of Vision, 11(14), 1–16, 10.1167/11.14.7. [DOI] [PubMed] [Google Scholar]
- Palmer, J. (1994). Set-size effects in visual search: The effect of attention is independent of the stimulus for simple tasks. Vision Research, 34(13), 1703–1721, 10.1016/0042-6989(94)90128-7. [DOI] [PubMed] [Google Scholar]
- Palmer, J. (1995). Attention in visual search: Distinguishing four causes of a set-size effect. Current Directions in Psychological Science, 4(4), 118–123, 10.1111/1467-8721.ep10772534. [DOI] [Google Scholar]
- Peirce, J., Gray, J. R., Simpson, S., MacAskill, M., Höchenberger, R., Sogo, H., … Lindeløv, J. K. (2019). Psychopy2: Experiments in behavior made easy. Behavior Research, 51(2), 195–203, 10.3758/s13428-018-01193-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- R Core Team. (2024). R: A Language and Environment for Statistical Computing. Vienna, Austria: R Foundation for Statistical Computing, https://www.R-project.org/. [Google Scholar]
- Rosenholtz, R. (2016). Capabilities and limitations of peripheral vision. Annual Review of Vision Science, 2, 437–457, 10.1146/annurev-vision-082114-035733. [DOI] [PubMed] [Google Scholar]
- Rosenholtz, R., Li, Y., Mansfield, J., & Jin, Z. (2005). Feature congestion, a measure of display clutter. In Proceedings of the SIGCHI conference on human factors in computing systems, Portland: Oregon, 761–770). [Google Scholar]
- Rosenholtz, R., Li, Y., & Nakano, L. (2007). Measuring visual clutter. Journal of Vision, 7(2), 17.1–22, 10.1167/7.2.17. [DOI] [PubMed] [Google Scholar]
- Semizer, Y., & Michel, M. (2022). Target-specific image clutter metrics for visual search. Journal of Vision, 22(14), 4365. Retrieved from 10.1167/jov.22.14.4365. [DOI] [Google Scholar]
- Semizer, Y., & Michel, M. M. (2019). Natural image clutter degrades overt search performance independently of set size. Journal of Vision, 19(4), 1, 10.1167/19.4.1. [DOI] [PubMed] [Google Scholar]
- Semizer, Y., & Rosenholtz, R. (2025). The effect of background clutter on visual search in video conferencing. Cognitive Research: Principles and Implications, 10(1), 40. [DOI] [PMC free article] [PubMed] [Google Scholar]
- Simoncelli, E., & Freeman, W. (1995). The steerable pyramid: A flexible architecture for multiscale derivative computation. In Proceedings of the 2nd IEEE International Conference on Image Processing (Vol. 3, pp. 444–447). Washington, DC, 10.1109/ICIP.1995.537667. [DOI] [Google Scholar]
- Tufte, E. R. (2001). The visual display of quantitative information (2nd ed.). Cheshire, CT: Graphics Press. [Google Scholar]
- van den Berg, R., Cornelissen, F. W., & Roerdink, J. B. T. M. (2009). A crowding model of visual clutter. Journal of Vision, 9(4), 24.1–11, 10.1167/9.4.24. [DOI] [PubMed] [Google Scholar]
- Wolfe, J. (1998). Visual search. In Pashler H. (Ed.), Attention (pp. 13–71). London: University College London Press. [Google Scholar]
- Wolfe, J., Oliva, A., Horowitz, T. S., Butcher, S. J., & Bompas, A. (2002). Segmentation of objects from backgrounds in visual search tasks. Vision Research, 42(28), 2985–3004. 10.1016/S0042-6989(02)00297-1. [DOI] [PubMed] [Google Scholar]
- Yu, C.-P., Samaras, D., & Zelinsky, G. J. (2014). Modeling visual clutter perception using proto-object segmentation. Journal of Vision, 14(7), 1–16, 10.1167/14.7.4. [DOI] [PMC free article] [PubMed] [Google Scholar]





