Ordinal focal loss: A relative fashionability ranking learning method

Existing fashion recommendations often rely on natural language processing or content-based image retrieval, overlooking direct aesthetic assessments of fashion images. Given the subjectivity and complexity of this task, we propose treating fashionability as a relative attribute to rank paired clothing images. To address this ranking challenge, we propose Ordinal Focal Loss, which transforms the pairwise ranking problem into a multi-classification task, leveraging ordinal attributes to improve classification boundaries. Furthermore, in terms of fashion feature representation, we propose modeling not just individual items but also their combined effect as an outfit, providing a more holistic and nuanced fashion representation. We introduce the Fashionability3k dataset, comprising 3k image pairs (2398 ordered and 601 similar pairs) with objective relative fashion labels. Experiments on three datasets—our Fashionability3k and two public datasets—show that our method outperforms the baseline by nearly 1 % in ranking accuracy. In the user study, it achieved a 0.72 consistency with human subjective perception. Moreover, combining local and global visual features leads to additional performance gains, with an average improvement of 2.78 % in ordered pairs and 1.09 % in similar pairs. This is the first study to treat fashionability as an objective attribute for comparative analysis, validated through extensive experiments.