PaperlayerView exact revision on arXiv

Source-linked AI summary

SCUT-FBP: A Benchmark Dataset for Facial Beauty Perception

Duorui Xie, Lingyu Liang, Lianwen Jin, Jie Xu, Mengru Li

arXiv:1511.02459v1cs.CV

TL;DR

Face databases used for attractiveness prediction may lack attractiveness ratings, and existing beauty datasets have limitations for facial beauty analysis. The paper develops SCUT-FBP, a dataset of standardized Asian female portraits with attractiveness ratings, and evaluates geometric and texture features using traditional machine learning and deep learning. The best Pearson correlation was 0.8187 for deep learning, versus 0.6482 for traditional machine learning.

  • Problem

    Face databases used for attractiveness prediction may lack attractiveness ratings, and existing beauty datasets have limitations for facial beauty analysis.

    Source: §I · p. 1

  • Method

    The paper develops SCUT-FBP, a dataset of standardized Asian female portraits with attractiveness ratings, and evaluates geometric and texture features using traditional machine learning and deep learning.

    Source: §I · p. 2

  • Results

    The best Pearson correlation was 0.8187 for deep learning, versus 0.6482 for traditional machine learning.

    Source: §I · p. 2

  • Takeaways & Limitations

    The authors conclude that SCUT-FBP provides a reliable benchmark for facial beauty analysis.

    Source: §I · p. 2, §V · p. 6

Abstract

from arXiv · show

In this paper, a novel face dataset with attractiveness ratings, namely, the SCUT-FBP dataset, is developed for automatic facial beauty perception. This dataset provides a benchmark to evaluate the performance of different methods for facial attractiveness prediction, including the state-of-the-art deep learning method. The SCUT-FBP dataset contains face portraits of 500 Asian female subjects with attractiveness ratings, all of which have been verified in terms of rating distribution, standard deviation, consistency, and self-consistency. Benchmark evaluations for facial attractiveness prediction were performed with different combinations of facial geometrical features and texture features using classical statistical learning methods and the deep learning method. The best Pearson correlation (0.8187) was achieved by the CNN model. Thus, the results of our experiments indicate that the SCUT-FBP dataset provides a reliable benchmark for facial beauty perception.

I. INTRODUCTION

The paper addresses the shortage of publicly available facial-beauty datasets with suitable attractiveness ratings by introducing SCUT-FBP as a benchmark. It also evaluates feature-based and deep-learning approaches, with the CNN achieving the highest reported Pearson correlation.

  • Existing face databases used for attractiveness prediction may lack attractiveness ratings, while available beauty datasets have limits for facial structure, image quality, or face-specific assessment.Source: §I · p. 1
  • SCUT-FBP provides high-resolution, front-on portraits of Asian female subjects with neutral expressions, simple backgrounds, and minimal occlusion, alongside ratings verified across four aspects.The average number of raters per image is 70.Source: §I · p. 2
  • The benchmark evaluates an 18-dimensional geometry feature and 2-dimensional Gabor texture features, as well as traditional machine-learning and deep-learning methods.Source: §I · p. 2
  • Pearson correlation was 0.8187 for deep learning and 0.6482 for traditional machine learning, supporting SCUT-FBP as a benchmark for facial beauty analysis.Source: §I · p. 2

II. CREATION OF SCUT-FBP

SCUT-FBP comprises 500 portraits collected from multiple sources and rated by numerous raters. The dataset uses a standardized portrait format and includes a higher proportion of beautiful faces than the general population to facilitate learning.

A. Data collection

  • SCUT-FBP contains 500 portraits, some captured by the authors and others licensed or downloaded from the Internet.Source: A. Data collection · p. 2
  • The portraits are high-resolution, front-on images of Asian female subjects with neutral expressions, simple backgrounds, no accessories, and minimal occlusion.The authors say this standardized format reduces effects of factors such as age, gender, and facial expression.Source: A. Data collection · p. 2
  • The dataset has a higher proportion of beautiful faces than the general population to facilitate effective learning of facial beauty.Source: A. Data collection · p. 2

B. Rating collection

Ratings were collected through a web-based assessment system used by 75 raters, with an average of 70 raters per portrait. Ratings were analyzed to define each portrait’s attractiveness label, and a subset of raters re-rated portraits to assess self-consistency.

B. Rating collection

  • 75 raters used the web-based facial beauty assessment system, with an average of 70 raters per image.Portraits were shown randomly, and raters could change a previous rating.Source: B. Rating collection · pp. 2–3, Figure 2 · p. 2
  • The authors analyzed ratings, selected appropriate data, omitted erroneous data, and defined each portrait’s attractiveness label as the average rating.Source: B. Rating collection · p. 3
  • Twenty original raters—10 male and 10 female—re-rated portraits after one and two weeks to verify self-consistency.Source: B. Rating collection · p. 3

III. ANALYSIS OF SCUT-FBP

The dataset analysis covers rating distribution, standard deviation, consistency, and self-consistency.

III. ANALYSIS OF SCUT-FBP

  • The analysis examines rating distribution, standard deviation, consistency, and self-consistency.Source: §III · p. 3

A. Rating distribution

The rating distribution is nearly Gaussian, with most portraits averaging around 2.5 and a smaller peak near 4.5. Rating dispersion varies with mean rating, with average-rated portraits showing greater disagreement.

A. Rating distribution

  • Most portraits average around 2.5 in the nearly Gaussian rating distribution, while a smaller peak near 4.5 reflects deliberate oversampling of beautiful faces.The authors say the higher proportion of beautiful faces was included to facilitate learning.Source: A. Rating distribution · p. 3, Figure 4 · p. 3

B. Standard deviation

  • Portraits rated between 2.5 and 3.5 tend to have higher rating standard deviations than portraits rated below 2.5 or above 3.5.Standard deviation decreases as ratings approach 1 or 5.Source: B. Standard deviation · p. 3, Figure 5 · p. 3
  • Ratings indicate more unified opinions about beautiful faces, whereas perceptions of average faces are more subjective.Source: B. Standard deviation · pp. 3–4

C. Consistency

The authors assess rating consistency through split-group comparisons, a t-test, and checks across rating sets collected at different times.

C. Consistency

  • 0.966–0.973 was the correlation between two group mean ratings, exceeding the previously reported ranges of 0.9–0.95 and 0.87–0.9.Source: C. Consistency · p. 4
  • The mean ratings of the two groups were not statistically different in the t-test.Source: C. Consistency · p. 4

D. Self-consistency

IV. BENCHMARK EVALUATION

The benchmark evaluation treats facial attractiveness rating as regression and compares traditional machine-learning with deep-learning methods.

  • Facial attractiveness rating is treated as a regression problem, with benchmark evaluations comparing traditional machine learning and deep learning.Source: §IV · p. 4

A. Facial beauty prediction using traditional machine

Traditional prediction experiments evaluate geometric, texture, and combined features with statistical learning algorithms. The combined geometric and UniSampleGabor features perform better than either feature type alone.

a) Geometric features

b) Skin texture features

2) Facial beauty prediction

a) Performance based on geometric features

b) Performance based on texture features

c) Performance based on combination of texture and geometric features

B. Facial beauty prediction using deep learning

The section contrasts automatic feature learning with manual feature extraction and evaluates a CNN for facial beauty prediction. Across five random splits, the CNN achieved an average correlation coefficient of 0.8187.

B. Facial beauty prediction using deep learning

V. CONCLUSION

The paper introduces the SCUT-FBP dataset and reports that its best Pearson correlation was 0.8187 with the CNN model. Its ratings were analyzed for reliability, and the dataset is presented as a resource for further facial beauty analysis.

V. CONCLUSION

  • 0.8187 was the best Pearson correlation, achieved by the CNN model.Source: §V · p. 6
  • SCUT-FBP contains attractiveness-rated portraits of 500 Asian female subjects and is publicly available.Source: §V · p. 6
  • The authors analyzed ratings from multiple aspects, confirming dataset reliability; they propose it for investigating facial beauty analysis problems.Source: §V · p. 6
Loading 1511.02459v1…