Designovel's system description for Fashion-IQ challenge 2019

Abstract: This paper describes Designovel's systems which are submitted to the Fashion IQ Challenge 2019. Goal of the challenge is building an image retrieval system where input query is a candidate image plus two text phrases describe user's feedback about visual differences between the candidate image and the search target. We built the systems by combining methods from recent work on deep metric learning, multi-modal retrieval and natual language processing. First, we encode both candidate and target images with CNNs into high-level representations, and encode text descriptions to a single text vector using Transformer-based encoder. Then we compose candidate image vector and text representation into a single vector which is exptected to be biased toward target image vector. Finally, we compute cosine similarities between composed vector and encoded vectors of whole dataset, and rank them in desceding order to get ranked list. We experimented with Fashion IQ 2019 dataset in various settings of hyperparameters, achieved 39.12% average recall by a single model and 43.67% average recall by an ensemble of 16 models on test dataset.