Comparative Analysis of Visual Attribute Tagging Models for Upper-Body Clothing Products

Accurate classification of fashion attributes is crucial for improving product recommendations and customer experience in e-commerce. In this study, we evaluate the performance of three advanced classification models, EfficientNet, Vision Transformer, and CatBoost Classifier with FashionClip embeddings, on various fashion attribute classification tasks, which including sleeve length, pattern type, neckline, and fit. We use upper-body clothing products from LC Waikiki catalogue. Our experimental results indicate that all models achieve high recognition rate in sleeve length classification, with accuracy of 0.97. The CatBoost model demonstrates superior performance in pattern type classifications with accuracy score of 0.88, highlighting its robustness in capturing intricate patterns. EfficientNet excels in neckline classification with an accuracy score of 0.91, particularly for round neck and shirt collar types, while Vision Transformer shows competitive but slightly lower performance. Fit type proved to be the most challenging attribute to classify accurately from product images. These findings highlight the strengths and limitations of each model, offering valuable insights for future enhancements in fashion attribute classification.