Personalized Image Aesthetics Assessment with Rich Attributes

1. Introduction

• PARA adds content preference, judgment difficulty, emotion, and willingness to share alongside image-oriented attributes to characterize personalized preferences.The paper also provides desensitized subject information, including age, gender, education, personality traits, and photographic and artistic experience.Source: §1 · p. 2
• 31,220 images annotated by 438 subjects form PARA, with 4 human-oriented subjective attributes, 9 image-oriented objective attributes, and desensitized subject information.Each image received annotations from an average of 25 subjects.Source: §1 · p. 2
• Statistical analysis indicates that personality traits, judgment difficulty, and image emotion can mirror personalized aesthetic preferences.The authors suggest these relationships as research angles for modeling preferences with subject information.Source: §1 · p. 2
• Training with human-oriented annotations as subject information improves PIAA model performance in the PARA benchmark.The benchmark compares unconditional and conditional PIAA models.Source: §1 · p. 2

2.1. Databases for PIAA

• FLICKR-AES, REAL-CUR, and AADB provide different PIAA resources, but vary in scale, annotation coverage, and intended use.FLICKR-AES has 40,000 images rated by 210 workers; REAL-CUR contains 14 personal albums and is usually used for testing; AADB has 10,000 images rated by 190 workers and provides 11 aesthetics attributes.Source: §2.1 · pp. 2–3

2.2. PIAA Models

• Prior PIAA models use GIAA priors, multimodal collaborative learning, multi-task learning, and meta-learning to capture personalized preferences.The review notes that these approaches introduce extra information into learning, including user-specific data and personality information.Source: §2.2 · p. 3

3. PARA Database

• PARA’s construction comprises four stages: data collection, label-system design, subject selection, and subjective experiments.Source: §3 · p. 3

3.1. Data Collection

• Images were collected from Creative Commons-licensed Flickr sources, scene-labeled and manually checked, then supplemented to balance scene diversity and aesthetics scores.Around 28,000 images were sampled by scene, with around 3,000 more from Unsplash and image-quality databases including SPAQ and KonIQ-10K.Source: §3.1 · pp. 3–4

3.2. Label system design

• Each image receives 13 labels alongside subject information, spanning image-oriented scores and overall quality, scene category, and aesthetics score.Scene categories are pre-annotated; the aesthetics score ranges from 1 to 5, with a middle choice between integer levels.Source: §3.2 · p. 4
• The label system also captures image emotion, judgment difficulty, content preference, willingness to share, and subject demographics and experience.Emotion is a single dominant choice; difficulty uses -1, 0, or +1, while content preference and sharing willingness use five-point scales.Source: §3.2 · p. 4

3.3. Subject Selection

• Subjects were selected for annotation experience, portrait diversity, personality-trait coverage, good health, and passing a color-blindness test.Subjects may request deletion of their annotations and leave the experiment.Source: §3.3 · pp. 4–5

3.4. Subjective Experiments

• The experiments used 446 sessions with unlabeled, standard, and repeated images to check annotation quality and consistency.The web tool required subjects to complete all fields before submission and allowed them to revise annotations.Source: §3.4 · p. 5

4.1. Data Summary

• PARA contains 31,220 images rated by 438 subjects, with nine scene categories pre-balanced at nearly 10% each.Source: §4.1 · p. 5

4.2. Statistical Analysis

• In (4, 5], aesthetic-score variance is lower than in other intervals, while attribute distributions are similar but slightly different.Source: Figure 4 · p. 5, §4.2 · p. 5, Figure 5 · p. 6
• Aesthetics and image quality are extremely highly correlated; most attribute correlations are around 0.5, while content preference and willingness to share are also highly correlated.Source: Figure 6 · p. 6, §4.2 · p. 6

4.3. Subject preference

• Three subjects with different personality traits rated one image 4, 2.5, and 3.5; high-N subjects diverged, while high-E subjects' content-preference correlation exceeded 0.8.Source: §4.3 · p. 6, Figure 7 · p. 6
• Subjects find low-aesthetics photos harder to judge, citing dizziness, unclear scenes, and difficulty distinguishing “worse” from “much worse.”Source: §4.3 · pp. 6–7, Figure 9 · p. 7
• Images scoring below 2.0 tend to convey negative emotion, while those scoring above 4.0 tend to convey positive emotion.Source: Figure 10 · p. 7, §4.3 · p. 7

5.1. Conditional and Unconditional PIAA

• The benchmark compares unconditional GIAA-to-PIAA fine-tuning with conditional models using personality, artistic experience, and photographic experience, across no-, 10-, and 100-shot settings.The unconditional ResNet-50 results are designated the official PARA benchmark.Source: §5.1 · p. 7, Table 3 · p. 8

5.3. Experimental Results

• Both 10- and 100-shot fine-tuning outperform the control group; more personal data improves fine-tuning, and subjective attributes improve PIAA over unconditional modeling.Source: Table 3 · p. 8, §5.3 · p. 8

5.2. Evaluation Criteria

• GIAA is evaluated with MSE, SROCC, PLCC, and classification accuracy, while PIAA uses SROCC and PLCC.Source: §5.2 · p. 8

5.3. Experimental Results

• The study tests five GIAA backbones, with their backbone comparison reported in Table 4.Source: §5.3 · p. 8, Table 4 · p. 8

• PARA contains 31,220 images annotated by 438 subjects, with nine image-oriented attributes, four human-oriented attributes, and desensitized subject information.Source: §6 · p. 8
• Statistical results indicate that personalized aesthetic preferences can be mirrored by human-oriented subjective attributes.Source: §6 · p. 8
• Adding subject information as a conditional prior improves modeling of personal aesthetic preference in the reported experiments.Source: §6 · p. 8