Leveraging Adult Fashion to Enhance Children’s Fashion Recognition

Abstract: The future of the fashion industry is expected to be online, thus a significant amount of research is being conducted in the field of fashion image analysis. Currently, a task that places a heavy workload on online stores is manually tagging new garments, including attributes such as category, color, pattern, or style. To this end, extensive research has targeted the automatic prediction of clothing categories and attributes, achieving promising results. Nevertheless, no previous study has been found in the literature that specifically reflects the performance of clothing attribute recognition with children’s clothing. This work intends to fill this gap and effectively present, in the same fashion analysis task, how a model trained in adult fashion performs over a model trained exclusively in children’s fashion. When examining the global understanding of children’s fashion apparel, the experiments exhibit that the best performance is obtained when leveraging the domain knowledge of adult fashion, specifically from the iMaterialist dataset, wherein the best model a difference in the overall performance of about 3% was achieved compared to pre- training on the ImageNet dataset or 12% when only children’s fashion was considered for training.