Automatic Spatially-aware Fashion Concept Discovery

[jchanxtarov]

Basic Information

• Authors: Xintong Han, Zuxuan Wu, Phoenix Huang, Xiao Zhang
• Date: Aug, 2017
• Published By: arXiv

Link

https://arxiv.org/pdf/1708.01311.pdf

Overview

• アイテム画像と単語（商品説明）を同一空間に埋め込む（image(pink skirt) - pint + blue = blue skirt）
• カテゴリごとに個別の空間に埋め込むので，モデルは個別に学習
• Stochastic bidirectional contrastive loss (to achieve good convergence)
  • contrastive loss: min( max( 0, hyperparam - cosine_dis(pos_img_embed, pos_rev_embed) + cosine_dis(pos_img_embed, neg_rev_embed) ) + max( 0, hyperparam - cosine_dis(pos_rev_embed, pos_img_embed) + cosine_dis(pos_rev_embed, neg_img_embed) ) )
• pre-trained CNN(GoogleNet, GAP layer) -> 2048 ... (面白いので別途解説を書く)

Others

Reference (for understanding)

[jchanxtarov]

実装する