Core Concepts
A novel end-to-end facial attractiveness prediction approach that integrates dual label distribution and lightweight design to achieve promising results with greatly decreased parameters and computations.
Abstract
The paper presents a novel end-to-end facial attractiveness prediction (FAP) approach that integrates dual label distribution and lightweight design. The key highlights are:
- Dual Label Distribution:
- The manual ratings, attractiveness score, and standard deviation are aggregated explicitly to construct a dual-label distribution, including the attractiveness distribution and the rating distribution.
- The attractiveness distribution is generated using the Laplace distribution, while the rating distribution is derived directly from the manual ratings.
- The dual-label distribution is designed to make the best use of the dataset.
- Joint Learning Framework:
- The dual-label distribution and the attractiveness score are jointly optimized under three learning modules: attractiveness distribution learning, rating distribution learning, and score regression learning.
- The joint learning framework aims to leverage the complementary information in the dual-label distribution to enhance the prediction performance.
- Lightweight Design:
- The data preprocessing and augmentation are simplified to a minimum.
- MobileNetV2 is selected as the backbone to balance performance and efficiency.
Extensive experiments on two benchmark datasets demonstrate that the proposed approach achieves state-of-the-art or comparable results while greatly reducing the number of parameters and computations compared to previous methods. Ablation studies and visualization further validate the effectiveness of the delicately designed learning modules and the model's ability to perceive facial attractiveness.
Stats
The average attractiveness score ranges from 1 to 5.
The standard deviation of the attractiveness scores is provided in the dataset.
The number of manual ratings for each image is also available.
Quotes
"Facial attractiveness plays a significant role in daily life [1], [2]. It is a complex and multifactorial concept, devoting researchers from diverse disciplines to decrypting its mysteries [3]."
"With the emergence of deep learning, convolutional neural networks (CNNs) [14]–[16] have been applied to FAP. Due to their powerful nonlinearity, CNN-based methods are able to learn hierarchical aesthetic representations thus boosting the performance, but their large-scale models lack flexibility."