AI algorithms, and in particular Machine Learning (ML) algorithms, learn from data tasks that have been traditionally done by humans such as: image classification, facial recognition, linguistic translation etc. To have a good generalization capability, AI algorithms must learn from sufficiently representative data, which is unfortunately not often the case. This results in a hyper-specialization of AI and its inability to perform well on new data whose distribution is too far from the one of the training set. It raises ethical questions which will undoubtedly have direct or indirect consequences on society. However, and despite biases they can entail, AI technologies are revolutionizing virtually every industry, and are forcing players in those industries to reinvent their businesses.
Like many industries, the fashion industy is being hit hard by exponential advances in AI. AI technologies are now able to describe a dress model, extract its attributes (color, style, type of sleeve, etc.), predict future fashion trends and even offer personalized clothing styles. Several publicly large datasets of fashion images made it possible.
However, the lack of diversity in available datasets is palpable. For example, images of African fashion are almost absent from these datasets. It raises problems in terms of ability to generalize algorithms trained on those datasets to African styles for instance, and therefore it limits the adoption of AI technologies within the African fashion industry.
Therefore, for an inclusive AI in the field of fashion, and to ensure that African fashion can benefit from the potentials of AI, Ai4Innov initiated the Inclusive Fashion AI project (InFashAI) which aims to create datasets that are much more representative of the diversity that exists in the world of fashion. We will first focus on building up a large volume of data on African fashion. This dataset will be progressively open source and, we hope, will be the backbone for AI tools adapted to African fashion.
We are releasing the first version the dataset named InFashAIv1. This dataset contains 15,716 images. You can download images from this Google drive repository. The dataframe containing titles, prices and descriptions is on this Github repository under the name infashaiv1.csv. It is a CSV file with an # separator.
@article{hacheme2021neural, title={Neural Fashion Image Captioning: Accounting for Data Diversity}, author={Hacheme, Gilles and Sayouti, Noureini}, journal={arXiv preprint arXiv:2106.12154}, year={2021} }