Kaleido-BERT: Vision-Language Pre-Training on Fashion Domain

http://dpfan.net/Kaleido-BERT Figure 1: Vision-Language (VL) pre-training architecture on fashion. We propose a novel VL pre-training architecture (Kaleido-BERT), which consists of a Kaleido Patch Generator (KPG), Attention-based Alignment Generator (AAG), and Alignment Guided Masking (AGM) strategy to learn better VL feature embeddings. Kaleido-BERT achieves the state-ofthe-art on the standard public Fashion-Gen dataset and deploys to the online system (a).