Fashionpedia-Taste: A Dataset towards Explaining Human Fashion Taste (2305.02307v1)

Abstract: Existing fashion datasets do not consider the multi-facts that cause a consumer to like or dislike a fashion image. Even two consumers like a same fashion image, they could like this image for total different reasons. In this paper, we study the reason why a consumer like a certain fashion image. Towards this goal, we introduce an interpretability dataset, Fashionpedia-taste, consist of rich annotation to explain why a subject like or dislike a fashion image from the following 3 perspectives: 1) localized attributes; 2) human attention; 3) caption. Furthermore, subjects are asked to provide their personal attributes and preference on fashion, such as personality and preferred fashion brands. Our dataset makes it possible for researchers to build computational models to fully understand and interpret human fashion taste from different humanistic perspectives and modalities.

• The paper presents a novel dataset integrating localized attributes, human attention, and caption annotations to elucidate fashion taste.
• The dataset comprises 10,000 expressions, 52,962 attribute selections, 20,000 attention markers, and 20,000 captions from 100 annotators across 1500 images.
• The study demonstrates its potential to enhance AI-driven fashion recommendation systems by providing interpretable insights into consumer behavior.

Fashionpedia-Taste: A Dataset towards Explaining Human Fashion Taste

The paper "Fashionpedia-Taste: A Dataset towards Explaining Human Fashion Taste," presents a novel approach to understanding consumer preferences in the fashion domain through the introduction of the Fashionpedia-Taste dataset. This dataset aims to enhance the interpretability of fashion-related liking/disliking behavior by encompassing multiple humanistic perspectives. The work proposes a framework that challenges existing computer vision systems to not only predict human preferences for fashion images but also to elucidate the underlying rationale behind these preferences.

Summary of the Dataset

Fashionpedia-Taste comprises data annotated from three distinct perspectives to explain reasons for liking or disliking fashion images:

1. Localized Attributes: Annotating fashion items based on specific attributes, thus allowing nuanced insights into what contributes to their appeal or detraction.
2. Human Attention: Capturing regions of images that draw particular focus, simulating human gaze and interest.
3. Caption: Providing textual descriptions that complement human attention data, offering verbal explanations for specific interests in an image.

In addition, the dataset collects personal attributes and preferences, such as personality traits and favorite brands, offering a multi-faceted view of human-fashion interaction. The dataset comprises 10,000 expressions, 52,962 attribute selections, 20,000 human attentions, and 20,000 captions, all annotated by 100 unique subjects across over 1500 unique images.

Implications and Novelty

The research underscores a critical gap in previous studies: existing datasets often omit the reasoning context of consumer preferences. By explaining these preferences, Fashionpedia-Taste aims to refine recommendation systems, leading to deeper insights into user behavior in the fashion e-commerce sector. The dataset is distinct in integrating interpretability into fashion preference analysis, allowing models to fulfill explainability requirements in addition to basic liking predictions.

Fashionpedia-Taste highlights that two individuals can appreciate the same clothing item for entirely different reasons, revealing the importance of diverse perspectives in preference interpretation. This insight sets a unique precedent in dataset creation, potentially advancing the development of more nuanced AI systems in fashion analytics and beyond.

Theoretical and Practical Application

Theoretically, this work contributes to interpretability research, a burgeoning area crucial for the advancement of AI model efficacy and user trust. Practically, it proposes a new paradigm for fashion recommendation systems, capable of explaining predictions and thus improving consumer engagement and satisfaction.

The implications of the dataset stretch into personalized shopping experiences and tailored advertisement strategies, where deeper understanding of consumer tastes can lead to improved sales conversions and user experience.

Conclusion and Future Directions

Fashionpedia-Taste opens exciting avenues for future exploration in AI-driven fashion analytics. The dataset's ability to capture multipart reasoning through rich annotations fosters richer interaction models. Future research might build on this work by exploring cross-cultural fashion preferences or expanding the methodology to other consumer-facing digital sectors.

In conclusion, Fashionpedia-Taste confronts a notable challenge in AI interpretability within fashion contexts, providing a resource that could bridge the gap between machine perception and human preference understanding.