Attentive Hierarchical Label Sharing for Enhanced Garment and Attribute Classification of Fashion Imagery

Fine-grained information extraction from fashion imagery is a challenging task due to the inherent diversity and complexity of fashion categories and attributes. Additionally, fashion imagery often depict multiple items while fashion items tend to follow hierarchical relations among various object types, categories and attributes. In this study, we address both issues with a 2-step hierarchical deep
learning pipeline consisting of (1) a low granularity object type detection module (upper-body, lower-body, full-body, footwear) and (2) two classification modules for garment categories and attributes based on the outcome of the first step. For the category and attribute-level classification stages we examine a hierarchical label sharing (HLS) technique in two settings: (1) single-task learning (STL w/ HLS) and (2) multi-task learning with RNN and visual attention (MTL w/ RNN+VA). Our approach enables progressively focusing on appropriately detailed features for automatically learning the hierarchical relations of fashion and enabling predictions
on images with complete outfits. Empirically, STL w/ HLS reached 93.99% top-3 accuracy while MTL w/ RNN+VA reached 97.57% top-5 accuracy for category classification on the DeepFashion benchmark, surpassing the current state-of-the-art without requiring landmark or mask annotations nor specialised domain expertise.