Fashion Compatibility Learning Via Triplet-Swin Transformer

Owing to the rising standard of living, personal appearance, mainly matching clothes, is essential to people. Because the right clothes not only can beautify their appearance directly but can also increase their self-confidence. This study aims to help users find a matching pair of clothes by considering the intricate details to choose suitable and compatible clothes. In this paper, increasing the efficiency of feature extraction is very important because fashion has a complicated concept, and the extraction of features such as the overall shape, design, and texture of clothes can significantly impact understanding and learning the compatibility of clothes. Therefore, suitable global features can help a lot in understanding the compatibility of clothes. Transformers can extract global features better than convolution networks. We use Swin Transformer networks to extract the image features. We have trained a Triplet-Swin network to learn fashion compatibility, which achieves better accuracy than previous methods. We evaluated our model with AUC and FITB metrics and the Polyvore Outfit dataset.