Afro SpecDetect: A multimodal Transformer-based attributes retrieval system for African fashion images
Keywords: African Fashion, Captionning, Dataset
TL;DR: In this paper, we explore the innovative use of image typology and data annotation to enhance automatic captioning, revealing significant improvements in caption quality and precision
Abstract: Despite advancements in Artificial Intelligence for fashion, African fashion remains underrepresented. This paper presents Afro SpecDetect, a dataset for African fashion multimodal captioning, incorporating attributes like color, material, and fabric. Experiments demonstrate enhanced performance in Bleu and F1 scores when the items' typologies are provided as a context in addition to the image.
Submission Category: Machine learning algorithms
Submission Number: 83
Loading