BigFashion: A Large-Scale Dataset for Fine-Grained Attributes Recognition

The digital technology of clothing has become a new direction in the clothing industry. How to capture the attributes of clothing is the prerequisite for ‘the Internet of Clothes’. Clothing digitization also plays an important role in clothing care, matching, and recommendation, which is essential to the health and well-being of humans. Clothing attributes are rich and varied, and visual recognition is the most direct way. In recent years, with the development of large models, large-scale datasets have been released rapidly in generic fields. Unfortunately, in the clothing domain, there is a shortage of large-scale datasets. In this paper, we construct a hierarchical clothing label attribute system, which describes a piece of clothing from multiple perspectives and dimension. Moreover, we introduce the BigFashion dataset, which is the largest dataset for clothing fine-grained recognition as we know. It contains more than 1.5 million images and 229 sub-labels. Compared with existing clothing datasets, it covers a wider range of scenes, such as indoor family scenes, wearing, hanging and tiled clothing images. Furthermore, we propose an advanced FTN (Fashion TransNet Recognition) network, which is based on feature fusion. In order to verify the effectiveness of our proposed method, we conduct a lot of experiments. We have verified the generalization of BigFashion in various clothing-relevant tasks such as clothing detection, clothing multi-label recognition, clothing retrieval, cross-modal clothing retrieval and clothing multi-modal generation. BigFashion is expected to extend more and more complex tasks in the clothing domain and its complex feature representation ability can become a new benchmark, contributing to the development of fine-grained and multi-dimensional visual analysis.