Eye-Tracking and Emotion-Based Evaluation of Wardrobe Front Colors and Textures in Bedroom Interiors

Abstract

Wardrobe fronts form a major visual element in bedroom interiors, yet material selection for their colors and textures often relies on intuition rather than evidence. This study develops a data-driven framework that links gaze behavior and affective responses to occupants’ preferences for wardrobe front materials. Forty adults evaluated color and texture swatches and rendered bedroom scenes while eye-tracking data capturing attraction, retention, and exploration were collected. Pairwise choices were modeled using a Bradley–Terry approach, and visual-attention features were integrated with emotion ratings to construct an interpretable attention index for predicting preferences. Results show that neutral light colors and structured wood-like textures consistently rank highest, with scene context reducing preference differences but not altering the order. Shorter time to first fixation and longer fixation duration were the strongest predictors of desirability, demonstrating the combined influence of rapid visual capture and sustained attention. Within the tested stimulus set and viewing conditions, the proposed pipeline yields consistent preference rankings and an interpretable attention-based score that supports evidence-informed shortlisting of wardrobe-front materials. The reported relationships between gaze, affect, and choice are associative and are intended to guide design decisions within the scope of the present experimental settings.

1. Introduction

Wardrobe fronts and other large cabinet surfaces significantly shape the visual and emotional experience of domestic interiors. Yet, in contemporary custom home-furnishing practice, color and texture selections are often guided by designer intuition or small expert panels, with limited input from user data. This ad hoc approach leads to inconsistent outcomes and poor reproducibility, hindering the advancement of human-centric, evidence-based interior design [1].

Existing research indicates that early visual attention and gaze dynamics are closely linked to choice behavior, where initial fixations can predict preferences in decision-making tasks [2,3]. Meanwhile, the Pleasure–Arousal–Dominance (PAD) framework offers a reliable basis for quantifying emotional responses to materials and environments [4,5]. Color preferences are known to be influenced by ecological and contextual factors [6,7], while material properties such as gloss and roughness also affect aesthetic and environmental appraisal [8]. However, two key gaps impede design application: first, eye-movement metrics are seldom integrated with affective measures in a unified analytic framework that links gaze behavior, PAD responses, and preference outcomes, while not claiming causal identification under the current experimental design; second, few studies translate experimental findings into reproducible rankings for real-world design alternatives, particularly in the context of custom furniture like wardrobes—frequently used elements in home environments.

Compared with prior interior eye-tracking work that typically reports gaze–preference correlations or single-model predictors, we contribute a design-ready pipeline that (a) produces calibrated preference strengths (Bradley–Terry) with uncertainty [9], (b) explains and generalizes those ranks through an interpretable gaze-based Integrated Attention Index (IAI) derived from predictive models, and (c) links attention mechanisms to affect (PAD) and Desire through multivariate association (CCA) and a parsimonious path model. For designers, this combination turns laboratory measurements into reproducible material shortlists and actionable heuristics, rather than standalone statistical effects.

By integrating eye-tracking and PAD measures within a unified framework, we aim to establish an interpretable, data-driven pipeline that characterizes the associations between visual attention, affective appraisal, and preference within controlled interior-design stimuli. Specifically, we examine whether eye-tracking features predict user preferences for wardrobe colors and textures, how PAD mediates this relationship, and how a data-driven pipeline can generate applicable design rankings.

Our methodology combines isolated evaluations of color and texture with scene-based selections reflecting real wardrobe layouts. Using pairwise choices, we compute Bradley–Terry preference strengths with bootstrap confidence intervals, while single-item eye tracking informs an attention index trained via leave-one-subject-out validation. We further investigate how scene context modulates baseline preferences and whether attentional and affective mechanisms explain these shifts [10,11].

This paper introduces a Gaze–Emotion–Preference (GEP) framework that unifies early and sustained viewing measures with PAD and choice data in a reproducible model. It provides convergent validation through BT rankings, an attention index, and an external questionnaire, and offers practical tools for palette screening, texture shortlisting, and scene composition. Ultimately, the framework connects data analytics with design practice, supporting scalable, emotion-informed material selection and evidence-based decision-making in custom home-furnishing design. Here, “preference” is operationalized by (i) forced-choice selections modeled via Bradley–Terry strengths and (ii) Desire ratings, while “emotion” is quantified using the PAD dimensions. We emphasize that the framework is evaluated on a discretized set of representative materials and controlled renderings; external validity to broader material catalogs and naturalistic bedroom environments requires additional validation.

The research framework is illustrated in Figure 1.

2. Materials and Methods

2.1. Extraction of Visual Physical Properties

This step ensures that all visual inputs are perceptually standardized and comparable across experimental conditions. We derived device-independent descriptors of the stimuli to support analysis and replication. Color swatches were converted to CIE L*a*b* coordinates to approximate perceptual uniformity [12]; texture patches were summarized with rotation-invariant statistics. All computations were scripted to ensure identical preprocessing across isolated items and their scene instantiations.

The CIE L*a*b* and texture descriptors are device-independent representations used for analysis. The visual presentation of rendered stimuli remains display-dependent. To reduce variability, all experiments were conducted on the same monitor with fixed brightness settings, and the rendering and export pipeline used a consistent sRGB color space with gamma 2.2. No hardware colorimeter calibration was performed; instead, presentation was controlled by using a single display device and fixed display settings across all sessions.

2.1.1. Color Data Extraction and Adaptive Palette Coverage

Color samples were measured in CIE L*a*b* coordinates under CIE standard illuminant D65 and a 2° standard observer. Measurements were obtained using an X-Rite i1Pro 2 spectrophotometer with a 45°/0° measurement geometry under controlled laboratory conditions. For each color swatch, three repeated measurements were taken from a central region of the surface to reduce edge effects and local surface non-uniformity.

Measurement replicates were screened using the CIEDE2000 color-difference metric (ΔE00). Replicates exceeding a tolerance of ΔE00 = 1.0 were re-acquired, and swatches exhibiting unstable measurements after repetition were excluded from further analysis [13,14].

To cover the perceptual color space without redundancy, the measured L*a*b* coordinates were clustered using perceptual distances computed by ΔE00. Cluster medoids were selected as representative palette entries for subsequent experimental stimuli.

2.1.2. Texture Feature Extraction

To characterize texture structure while minimizing chromatic influences, texture images were converted to grayscale prior to feature extraction. Rotation-averaged gray-level co-occurrence matrix (GLCM) statistics and local binary pattern (LBP) histograms were computed on these grayscale representations to capture spatial regularity and micro-pattern structure.

This processing choice isolates structural texture information from color cues in the feature space, ensuring that texture descriptors reflect surface pattern characteristics rather than chromatic variation.

2.1.3. Preparation of Interior-Scene Stimuli

Bedroom scenes were rendered with a fixed camera, neutral D65 illumination, and a standardized layout [15]. Regions of interest on wardrobe façades and background were pre-defined to map one-to-one to later eye-tracking AOIs [16,17]. Images were gamma-corrected, and background luminance was constrained to a narrow band to limit global-contrast confounds [18,19]. Elemental colors/textures were then embedded into the same ROIs with identical transforms, ensuring consistent interpretation between isolated items and their scene instantiations.

In both the element and scene conditions, AOIs were restricted to the wardrobe-front façade surfaces (door panels) and did not include adjacent walls, floor, ceiling, or decorative objects. Because the 3D model, camera pose, and wardrobe geometry were fixed, the AOI boundaries were identical across all stimuli within each condition; only the mapped color/texture of the wardrobe-front AOI changed. A separate background AOI (all non-wardrobe regions) was retained only for exposure normalization (e.g., fixation-share) and for data-quality checks.

2.2. Experimental

This section details participants, apparatus, and the unified procedure for the color–texture element task and the scene task. Unless noted, settings were identical across tasks.

2.2.1. Participants

Forty adults (20 female, 20 male; 20–30 years) took part. All reported normal or corrected-to-normal vision; myopic participants wore their habitual glasses. Individuals with diagnosed color-vision deficiency, clinically significant astigmatism, or large inter-ocular differences were excluded [20]. Written informed consent was obtained. Two trained researchers administered the sessions, one responsible for recording data and the other for operating the equipment. The study was conducted in accordance with institutional ethical guidelines and approved by the relevant ethics committee.

2.2.2. Experimental Equipment

The eye-tracking experiment utilized a Tobii Pro Fusion remote eye tracker, with a sampling rate of 250 Hz and a head movement range of 30 cm × 25 cm. Stimuli were presented on a Dell P2422H LCD monitor (23.8 inch, 1920 × 1080 resolution, 60 Hz refresh rate). Color management was handled at the operating-system level using the default sRGB ICC profile. The monitor was not hardware-calibrated with a colorimeter; however, all experiments were conducted on the same display with fixed brightness and contrast settings. The display was set to an sRGB color mode with a nominal D65 white point, gamma 2.2, and a target luminance verified at approximately 120 cd/m² using the built-in system calibration utility. These settings were held constant across all experimental sessions to minimize presentation variability. An MSI laptop was used for stimulus preparation, questionnaire presentation, and data export/inspection. Calibration followed the manufacturer’s multi-point routine and was repeated if accuracy fell below acceptable bounds [21].

Eye-movement events were parsed using a velocity-threshold identification method (I-VT) as implemented in the ErgoLAB–Tobii processing pipeline. A saccade velocity threshold of 30°/s was applied, and fixations shorter than 60 ms were excluded. All parameters were held constant across experimental conditions to ensure the comparability of fixation-based metrics, including time to first fixation, fixation count, mean fixation duration, scanpath length, and fixation distribution ratio.

Raw gaze samples marked as invalid by the Tobii SDK were removed prior to event detection. No participant-specific parameter tuning was performed.

2.2.3. Experimental Procedure and Task

Participants completed two forced-choice tasks in a counterbalanced order:

• Material-element task: On each trial, 6 swatches were presented. Participants indicated the most preferred option using a keypad or mouse;
• Bedroom-scene task: On each trial, 6 wardrobe scenes were presented. Participants indicated the most preferred configuration.

Each trial followed the same timing logic: central fixation 500 ms, stimulus display self-paced with a maximum of 6 s, response, and an inter-trial interval of 500–800 ms [22]. Before the formal session, participants completed 10–12 practice trials; successful calibration results are shown in Figure 2. To enable later fusion of attention and affect, brief PAD ratings were collected on nine-point scales using standardized SAM visuals [23]. The order of tasks and the spatial arrangement of alternatives were counterbalanced across participants to reduce position and sequence biases. The detailed experimental flow is illustrated in Figure 3.

The stimulus display was self-paced (maximum 6 s) to reduce time pressure and allow participants to complete multi-alternative comparisons. We note that self-paced exposure can couple viewing time with preference; therefore, fixation-duration-related predictors are interpreted as correlates of preference under this task design rather than causal determinants.

2.3. Data Acquisition and Analysis

The experiment comprised two choice tasks: a Material-element task (six isolated alternatives per trial) and a Bedroom-scene task (six rendered bedroom scenes per trial). For gaze-feature extraction, we additionally employed three single-element viewing conditions:

• Color-Single, presenting isolated color swatches;
• Color-Texture-Single, presenting the original texture samples;
• DeColor-Texture-Single, presenting the same texture samples after removing chromatic information to emphasize structural texture cues.

These single-element conditions were used to derive participant-level attention profiles and to construct interpretable gaze features for prediction.

All experimental records were consolidated into a long-format dataset indexed by participant, stimulus and task condition. Eye-tracking data from the three single-element tasks (Color-Single, DeColor-Texture-Single and Color-Texture-Single) were processed to extract five standardized metrics: fixation count (FC), time to first fixation (TTFF), mean fixation duration (MFD), saccade count (SC) and fixation-duration ratio (FDR) [24,25]. Time to first fixation was inverted so that higher values indicated faster orienting. Trials with missing or implausible gaze data were removed, and repeated presentations were averaged. Scene-choice data were encoded as binary labels for each trial. All features were z-scored using training-fold statistics to avoid information leakage [26].

2.3.1. Clustering of Visual-Attention Patterns

To characterize individual viewing strategies, each participant was represented by a vector containing the five gaze features concatenated across the three single-element tasks. These vectors were standardized and grouped using k-means clustering.

Cluster quality was evaluated using the Silhouette coefficient:

$$s(i)={\displaystyle \frac{b(i)-a(i)}{\max \{a(i),b(i)\}}},$$

(1)

where $a(i)$ is the mean intra-cluster distance for sample $i$ and $b(i)$ is the minimum mean inter-cluster distance, and the Calinski–Harabasz index:

$$\mathrm{CH}(k)={\displaystyle \frac{\mathrm{tr}(B_k)/(k-1)}{\mathrm{tr}(W_k)/(n-k)}},$$

(2)

where $B_k$ and $W_k$ denote between- and within-cluster dispersion matrices. Cluster membership was not used as a covariate in the main predictive models; instead, it served to ensure that model performance and feature-importance patterns remained qualitatively similar across attention-profile subgroups. Clustering was used as a descriptive check of heterogeneity in viewing strategies rather than as a primary explanatory variable. We report cluster sizes and verify that key modeling results are not driven by any single cluster through sensitivity analyses.

The clustering procedure identified three attention-profile clusters. Cluster sizes were n = 14, n = 13, and n = 13, respectively. The solution achieved a mean Silhouette coefficient of 0.41 and a Calinski–Harabasz index of 162.7, indicating moderate but stable separation among viewing-strategy groups.

Sensitivity analyses confirmed that removing any single cluster did not materially change the direction or magnitude of the main predictive and ranking results.

2.3.2. Preference Prediction Modeling

To assess whether gaze metrics predict material choice, each candidate in a trial was assigned a binary label ($y$ = 1 if chosen). A penalized logistic regression model was fitted using standardized FC, inverse TTFF, MFD, SC, FDR and, when available, relative fixation share. The choice probability was modeled as:

$$P(y=1|x)=\sigma ({\beta }_0+{\beta }^{T}x),\sigma (t)={\displaystyle \frac{1}{1+e^{-t}}},$$

(3)

with parameters estimated by maximizing the ${\mathcal{l}}_2$-regularized log-likelihood [27]

$$L(\beta )-\mathsf{\lambda }\left|\right|\beta |{|}_2^2.$$

(4)

Generalization was evaluated using leave-one-subject-out cross-validation. Model performance was summarized using AUC, accuracy, F1 score and top-1 hit rate [28]. A non-linear reference classifier was also fitted to assess robustness, and feature-importance patterns were compared qualitatively [29].

2.3.3. Bradley–Terry Pairwise Ranking

To obtain calibrated preference strengths, trial-level choices were transformed into winner–loser pairs and analyzed using the Bradley–Terry model [30]. Let ${\theta }_i$ denote the latent utility of stimulus $i$. The Bradley–Terry probability that stimulus $i$ is preferred over stimulus $j$ is given by

$$P(i\succ j)={\displaystyle \frac{\exp ({\theta }_i)}{\exp ({\theta }_i)+\exp ({\theta }_j)}}.$$

(5)

Utilities are identifiable up to an additive constant; for identifiability, we impose the constraint ${\displaystyle {\sum }_i{\theta }_i}=0$.

Each six-alternative trial was converted into winner–loser comparisons by pairing the selected option (winner) with each of the remaining five options (losers), yielding five comparisons per trial. Because these derived comparisons are not statistically independent within a trial, Bradley–Terry estimates are interpreted as aggregate preference strengths resulting from a multinomial-to-pairwise reduction. Uncertainty was therefore quantified using bootstrap resampling at the trial level.

Because each participant contributed multiple choices, we additionally conducted a robustness check using a mixed-effects Bradley–Terry specification with participant-level random intercepts. The resulting rank orders and top-tier materials were unchanged relative to the standard BT fit, indicating that clustering by observer did not materially affect the conclusions.

All robustness analyses were conducted in Python 3.13. Standard Bradley–Terry models were estimated using the choix package, and within-participant dependence was examined via mixed-effects logistic regression with participant-level random intercepts implemented in statsmodels. Rank-order correlations between the standard and mixed-effects models exceeded ρ = 0.95, and the top-ranked materials were identical across model specifications, indicating that participant-level clustering did not materially affect the preference conclusions.

Parameters were estimated using a majorization–minimization algorithm with a log-scale identifiability constraint. Bootstrap resampling provided percentile confidence intervals for ${\theta }_i$. Preference strengths were primarily estimated and reported separately for each task condition. Where a summary across conditions was reported, it was computed only for the common stimulus set shared across the corresponding tasks and was treated as a descriptive index rather than a standalone inferential estimate.

2.3.4. Integrated Attention Index

To produce an interpretable material score, feature weights were derived from the logistic model coefficients and normalized as

$$w_j={\displaystyle \frac{\left|{\beta }_j\right|}{{\displaystyle {\sum }_m\left|{\beta }_m\right|}}}$$

(6)

For each material $i$ the Integrated Attention Index (IAI) was computed as

$${\mathrm{IAI}}_i={\displaystyle \sum _j{w}_j{z}_{ij}}$$

(7)

where $z_{ij}$ is the standardized value of feature j [31,32].

The index provides a continuous ranking reflecting how perceptual behavior contributes to preference. When data coverage was insufficient for stable estimation, equal weights were used. Agreement with Bradley–Terry strengths and questionnaire results was used to assess consistency.

2.3.5. Linking Attention and Emotion

Eye-tracking features were aligned with Pleasure, Arousal and Dominance ratings for each participant–stimulus pair. Canonical correlation analysis quantified the shared structure between gaze variables X and PAD scores Y by maximizing

$$\rho =\underset{a,b}{\max }\mathrm{corr}({\mathrm{a}}^{\mathrm{T}}{\mathrm{X},\mathrm{b}}^{\mathrm{T}}\mathrm{Y})$$

(8)

revealing how attentional patterns co-vary with emotional appraisals [33].

Analyses were performed separately for color and texture datasets at the participant × stimulus level (color: n = 280 observations; texture: n = 240 observations). X comprised five standardized gaze variables (FC, TTFF−1, MFD, SC, FDR), and Y comprised three affective variables (Pleasure, Arousal, Dominance). Canonical correlations were tested using Wilks’ Lambda. The first canonical correlation was statistically significant in both datasets (p < 0.001), and subsequent dimensions were not interpreted. The path model was estimated as a standardized SEM (maximum likelihood with robust standard errors) with 5000 bootstrap resamples to obtain confidence intervals for key paths; all reported β values are standardized coefficients. Models were fitted in R (lavaan), and fit indices (CFI/TLI/RMSEA/SRMR) are reported in the Results.

For interpretability, three composite constructs—Attraction, Retention and Exploration—were created from standardized gaze features and entered into a sequential path model. In this model, the three constructs predicted Pleasure, Pleasure predicted Arousal and Arousal predicted Desire. Standardized path coefficients and bootstrap confidence intervals were used to evaluate direct and indirect effects. This integration provides a mechanistic bridge between visual behavior, affective response and material preference.

In this study, Desire is treated as an approach-oriented preference outcome reflecting willingness to select or adopt a material option. Desire is not a substitute for Dominance. PAD dimensions were collected and analyzed in the CCA stage; the subsequent path model focuses on a parsimonious Pleasure–Arousal–Desire sequence to connect affective appraisal to preference intention under our experimental setting.

3. Results

3.1. Stimulus Set

Materials were supplied by Shanghai Muli Industrial Development Co., Ltd. (Shanghai, China). Color swatches were transformed to CIE L*a*b* and clustered with a density-adaptive method; seven perceptually stable groups emerged and were represented by k-medoids, labeled by dominant hue as: Black, Grey, White, Wax white, Yellow, Red, and Green (Figure 4). Texture patches, summarized by rotation-averaged GLCM and LBP descriptors, formed six separable classes—Line Finish, Leather Finish, Mountain Grain, Straight Grain, Special Grain, and Horizontal Grain—with class medoids used as exemplars (Figure 5).

A standard linear wardrobe layout was adopted for presentation. A three-dimensional bedroom model was built in 3 ds Max. To limit extraneous influences, illumination was fixed to a uniformly bright level (≈90 ± 5 cd/m²), and decorative objects were omitted. All room elements other than the wardrobe were held constant. Wardrobe fronts were manipulated in two stimulus sets: a color set (levels as in Figure 4) and a texture set (categories as in Figure 5). Example stimuli are shown in Figure 6 and Figure 7. Because black exhibited very low fixation and choice rates in the single-color condition, it was excluded from the interior-scene color set. As a result, any cross-condition aggregation was limited to the common alternatives shared by the corresponding tasks to preserve comparability.

3.2. Eye-Tracking Preprocessing and Subject Stratification

Raw gaze streams were reduced to a long table keyed by subject–stimulus–condition with five metrics per item: fixation count (FC), time to first fixation (TTFF), mean fixation duration (MFD), saccade count (SC), and fixation-duration ratio (FDR). Latency was inverted to emphasize faster orienting. All variables were z-standardized; where relevant, we normalized exposure by the within-set fixation share. Trials with missing AOIs, zero dwell across items, or implausible latencies were excluded; repeated presentations were averaged.

For downstream analyses, we focused on single-element conditions (Color-Single; DeColor-Texture-Single; Color-Texture-Single). Each subject’s standardized features from these three conditions were concatenated to form an attention profile. Subjects were then stratified via k-means; the number of clusters was chosen by maximizing the mean of Silhouette and Calinski–Harabasz indices. Between-cluster differences in FC, TIFFinv, MFD, SC, and FDR were screened using Mann–Whitney tests with Holm correction. This yields a clean, comparable feature set and a subject stratification used in the subsequent prediction, ranking, weighting, and PAD-fusion analyses.

3.3. Preference Prediction Performance

Choices were modeled as binary outcomes from eye-tracking features aggregated at the subject–stimulus level. Six z-scored predictors were used: fixation count (zFC), inverse time-to-first-fixation (zTIFFinv), mean fixation duration (zMFD), saccade count (zSC), fixation-duration ratio (zFDR), and within-condition fixation share (zFR). Negatives were completed within each subject’s candidate set. Evaluation used leave-one-subject-out cross-validation. On each fold we trained L2-regularized logistic regression and a 200-tree random forest. We report AUC, accuracy, F1, Top-1 hit rate, standardized coefficients, and feature importances.

The results reveal a consistent landscape of gaze features across task conditions, as illustrated by the feature-mean heatmap (Figure 8). In the Color-Single condition, participants exhibited more active visual search, characterized by higher standardized fixation counts and saccade frequencies, together with shorter mean fixation durations. This pattern suggests rapid scanning behavior when evaluating isolated colors. In contrast, both Color-Texture-Single and DeColor-Texture-Single showed feature values close to zero on most axes, with modest increases in fixation-duration ratio, indicating more concentrated dwell within relevant regions during isolated texture appraisal. The Texture-Scene condition displayed a distinct profile marked by substantially longer mean fixation durations and markedly lower fixation-duration ratios, along with slightly reduced saccade counts. These patterns imply slower, more deliberate appraisal in the richer bedroom context, accompanied by more diffuse allocation of attention across the scene. Time to first fixation showed minimal variation across conditions, suggesting that the initial orienting response remained relatively stable despite differences in task structure and stimulus complexity.

3.4. Pairwise Preference Ranking

Bradley–Terry analyses of all pairwise choices yield clear, stable orders. Across isolated swatches and bedroom scenes, the rank order is White > Ivory white > Gray, with these achromatic options sitting above the normalized mean (geometric mean = 1) and showing limited overlap in 95% CIs. In scenes, strengths contract toward 1, indicating that context attenuates color contrasts while preserving the ordering (Figure 9a).

Mountain Grain ranks highest, followed by Straight Grain and Line Finish; this leading tier is separated from Leather Grain, Special Grain, and Horizontal Grain, which cluster around the mean with broader, overlapping CIs. Scene textures again compress toward the mean, mirroring the color pattern (Figure 9b).

Single-question checks reproduce the same structure: isolated color keeps White > Ivory white > Gray with wide separation, and isolated textures keep Mountain Grain > Straight Grain > Line Finish as the top tier; scene versions narrow the gaps. Taken together, achromatic colors and orderly/fine textures have the highest win probabilities, whereas contextual scenes systematically homogenize preferences.

3.5. Integrating Gaze Features for Preference Scoring

To explain BT ranks and project them to new designs, we quantified each eye-movement feature’s contribution to scene choices and combined them into an Integrated Attention Index (IAI). For color, mean fixation duration (MFD) dominates (44.0%), followed by inverse time-to-first-fixation (TIFFinv, 35.5%); saccade count (10.3%), fixation-duration ratio (6.9%), and fixation count (3.3%) add modest signal. For texture, MFD is even more dominant (62.6%), with TIFFinv (17.0%), SC (9.1%), FDR (9.4%), and FC (1.8%) contributing smaller increments. Thus, sustained viewing and rapid first looks carry most predictive power; simple counts contribute little.

Applying these weights to single-item sets yields IAI rankings that support and refine BT outcomes. For color, neutral lights lead—White first, then Ivory White and Gray—while saturated warms trail (

Table 1. IAI rankings for color stimuli.

Rank	Stimulus	IAI
1	White	1.349
2	Ivory White	1.339
3	Gray	0.010
4	Red	−0.059
5	Yellow	−0.073
6	Green	−0.144
7	Black	−0.183

). For texture, Mountain Grain, Straight Grain, and Line Finish score highest, with Leather Grain next and Special Grain lower (

Table 2. IAI rankings for texture stimuli.

Rank	Stimulus	IAI
1	Mountain Grain	1.180
2	Straight Grain	1.170
3	Line Finish	0.170
4	Leather Finish	0.078
5	Special Grain	−0.013
6	Horizontal Grain	−0.238

). Minor re-orderings versus BT align with the dominant role of MFD: items eliciting longer fixations and faster first looks gain IAI advantage when BT strengths are close.

The two analyses align well: the Bradley–Terry model establishes the preference hierarchy, and the Integrated Attention Index explains it through gaze behavior while enabling predictions for new materials. Options that attract faster initial fixations and sustain longer viewing receive higher IAI scores and are more likely to align with the top-performing neutral light colors and fine-grained textures. Conversely, materials that disperse attention or delay the first look tend to score lower and rank less favorably.

3.6. Linking Gaze to Affect

Eye-tracking metrics were standardized and organized into three attention constructs: Attraction, represented by inverse time to first fixation and fixation count; Retention, represented by mean fixation duration and the fixation-duration ratio; and Exploration, represented by saccade count. These constructs were first related to affective responses using canonical correlation analysis, and the resulting associations were then examined through a structural path model reflecting a sequential Pleasure–Arousal–Desire process.

Canonical correlation analysis revealed a significant association between gaze behavior and affective responses. The first canonical correlation was r = 0.62 for the color dataset and r = 0.58 for the texture dataset. Wilks’ Lambda tests indicated that the first canonical dimension was statistically significant in both cases (p < 0.001), whereas subsequent dimensions were not significant and were therefore not interpreted.

The proposed path model demonstrated an acceptable to good fit to the data. For the color dataset, model fit indices were CFI = 0.96, TLI = 0.94, RMSEA = 0.045, and SRMR = 0.038. Similar fit was observed for the texture dataset.

Standardized path coefficients from attention constructs to Pleasure and Desire were all statistically significant (p < 0.01). Bootstrap-based 95% confidence intervals confirmed the robustness of the estimated effects (

Table 3. Canonical correlation and path model summary.

Analysis	Statistic	Color	Texture
CCA	First canonical r	0.62	0.58
	Wilks’ Lambda (p)	<0.001	<0.001
SEM	CFI	0.96	0.95
	TLI	0.94	0.93
	RMSEA	0.045	0.048
	RMR	0.038	0.041

).

For both color and texture, the first canonical variate loaded positively on TTFF⁻¹, fixation count and FDR, and negatively on mean fixation duration. Thus, stimuli that are found quickly and receive dense early fixations align with higher PAD states, whereas prolonged viewing reflects a distinct retention process (Figure 10).

For color, attraction strongly predicts Pleasure (β ≈ 0.9), while Retention and Exploration are near zero. Pleasure in turn raises Arousal (β ≈ 1.1), and Arousal predicts Desire (β ≈ 0.7). Hence early attentional capture dominates color preference, with effects propagated along the Pleasure–Arousal–Desire pathway (Figure 11a).

For texture, a similar pathway emerged but with more differentiation among constructs. Attraction again strongly predicted Pleasure (β ≈ 1.0); Retention showed a small positive effect (≈0.1); Exploration is negative (≈−0.2 to −0.3). Pleasure again elevates Arousal (β ≈ 1.1), which increases Desire (β ≈ 0.65). Textures that concentrate attention early and suppress scanning obtain higher PAD and are chosen more often (Figure 11b).

These mechanisms help clarify the preference rankings. Neutral light colors such as white, ivory white and gray, along with the most orderly wood-grain textures, achieve higher desirability because they attract early and concentrated visual attention. When these materials are placed within full bedroom scenes, contextual factors reduce the magnitude of differences between options, yet the overall preference order remains unchanged.

3.7. Agreement with the Subjective Questionnaire

The independent questionnaire reproduced the BT hierarchy and IAI projections. For color, the top tier remained White, Ivory white, and Gray in both element and scene contexts; warmer chromatic hues consistently trailed. For texture, Mountain Grain, Straight Grain, and Line Finish formed the leading group, followed by Leather Finish, with Special Grain and Horizontal Grain near the middle-to-lower range. Minor reorderings were confined to mid-ranked items and did not alter tier boundaries. Overall, the survey yielded the same leaders and the same context-driven contraction observed in BT and IAI, supporting the external validity of the gaze–affect pipeline.

4. Discussion

4.1. Stable Preference Structures Across Interior Contexts

Across all analyses, wardrobe-front material preferences exhibited a clear and stable hierarchical structure that persisted from isolated swatches to full bedroom scenes. Neutral light colors—particularly white, ivory white, and gray—consistently occupied the top tier, while saturated hues and darker tones ranked lower. Similarly, structured wood-like textures such as mountain grain, straight grain, and line finish were preferred over more irregular or visually fragmented patterns.

When materials were embedded within realistic bedroom scenes, preference differences became less pronounced, yet the relative ordering remained unchanged. This pattern suggests a contextual compression effect rather than a reversal of preference: surrounding spatial elements, lighting, and furniture reduce contrast between options without altering their fundamental appeal. This finding is particularly relevant for interior practice, as it indicates that material rankings derived from controlled evaluations can remain informative when transferred to more complex spatial contexts.

4.2. Linking Visual Attention to Affective Evaluation and Choice

The eye-tracking results indicate that time to first fixation and mean fixation duration are the most influential predictors of wardrobe-front preference. Materials that attract attention rapidly and sustain longer visual engagement are more likely to be chosen, reflecting the combined importance of early perceptual salience and sustained evaluative processing.

By integrating PAD emotion ratings, the analysis further clarifies how attentional patterns translate into affective responses and ultimately into preference. Materials associated with higher Pleasure and Arousal tended to exhibit more efficient attentional profiles, suggesting a coherent mechanism linking perceptual processing, emotional appraisal, and choice behavior. This finding aligns with prior work in environmental psychology and decision science, while extending it to furniture-specific material evaluation in interior settings.

Importantly, the distinction between attraction, retention, and exploration highlights that not all gaze activity contributes equally to desirability. Rapid orienting and concentrated viewing appear to be more strongly associated with positive affect than extensive scanning, providing a more nuanced understanding of how visual behavior supports preference formation.

4.3. Implications for Bedroom Interior Design and Material Selection

The present findings are consistent with recent eye-tracking studies demonstrating that façade and interior surface attributes shape visual comfort and dwell patterns. At the same time, this study extends those insights by focusing on a decision-critical furniture component—wardrobe fronts—within controlled bedroom scenes, thereby bridging architectural-scale research and furniture-scale design decisions.

From a practical perspective, the results suggest clear design heuristics for bedroom interiors targeting young adult users. In compact or multipurpose bedrooms, combining neutral light-colored wardrobe fronts with orderly, fine-grained wood textures is likely to maximize perceived visual comfort and desirability. When materials are evaluated within full scenes, designers should expect preference differences to narrow rather than reverse; consequently, mid-ranked options may be differentiated more effectively through subtle adjustments in texture regularity, luminance, or surface finish rather than through large shifts in hue.

Beyond specific material recommendations, the integrated framework proposed in this study supports early-stage decision making in customized interior design. By combining calibrated pairwise rankings with an interpretable attention-based index, designers can efficiently screen large material sets and focus on options that consistently elicit favorable perceptual and affective responses.

4.4. Limitations and Directions for Future Research

Several limitations should be noted. The study examined a finite set of colors and textures, and the number of participants, although sufficient for controlled eye-tracking research, does not reflect the diversity of occupants encountered in real housing projects. Although grayscale descriptors reduce chromatic confounding in texture modeling, residual color cues in the original samples may still influence preference under real viewing conditions. All evaluations were conducted under laboratory conditions with standardized renderings and luminance, which differ from the varied lighting and furnishings of lived-in bedrooms.

Generalizability is therefore bounded: our participants were young adults (20–30 years), and material preferences can vary with age, culture, household composition, and life stage. Accordingly, the reported ranks and mechanisms should be interpreted as most representative of younger adult users under controlled viewing conditions, and future studies should replicate the pipeline with older and more diverse cohorts.

Because trials were self-paced and followed by repeated affect ratings, fatigue or attentional drift may influence gaze dynamics and judgment consistency. Future studies will therefore adopt fixed-duration presentations and block-wise temporal stability checks to disentangle preference effects from exposure-time endogeneity. In addition, alternative path specifications will be examined that explicitly incorporate the Dominance dimension, in order to test whether perceived control adds explanatory power for Desire across different contexts and populations.

More broadly, future work should expand the range of materials and include more diverse user groups to capture cultural, age-related, and lifestyle-specific differences in perception [34]. Field studies in furnished homes or immersive virtual reality environments may further clarify how dynamic lighting and spatial complexity shape attention and emotion. Beyond static gaze metrics, incorporating temporal transitions between fixations may yield richer indicators of perceptual strategies. Finally, embedding the proposed framework into digital design platforms could help demonstrate how data-driven evaluation supports large-scale customization and improves the consistency of interior material selection.

5. Conclusions

This study demonstrates a reproducible, data-driven approach to understanding how wardrobe front materials shape visual comfort and occupant preference in bedroom interiors. By integrating eye-tracking metrics with affective evaluations and calibrated pairwise modeling, the framework clarifies both the structure of material preferences and the mechanisms that drive them.

Three conclusions stand out. First, preference orders are stable across evaluation formats. Neutral light colors and structured wood-like textures consistently ranked highest, and scene context moderated but did not overturn these patterns. Second, visual attention plays a central role in material desirability. Rapid initial fixation combined with longer sustained viewing emerged as the most influential predictors of choice, linking perceptual efficiency with positive affective appraisal. Third, the integrated attention index provides a practical tool for screening new materials without requiring repeated pairwise tests, offering designers a scalable method for identifying visually comfortable and widely appealing wardrobe front options.

Overall, the findings contribute to human-centred interior design by offering quantitative evidence of how material attributes guide perception and preference. The proposed pipeline strengthens the connection between perceptual research and interior practice, supporting more consistent, evidence-based decision making in the design of bedroom environments.