Skip to content

Context-Aware Visual Compatibility Prediction #52

@kacky24

Description

@kacky24

short summary

GCNを、ファッションアイテム間のcompatibility予測に適用した論文。
単一のアイテム画像からの特徴量だけではなく、それに適合するアイテムを周辺ノードとしてエッジを貼ってGCNで考慮することで、精度を大きく向上。
スクリーンショット 2019-04-29 12 53 42

Encoderでは、GCNの層を積むことで特徴抽出を行うが、この際、ランダムにエッジをサンプリングすることで、構造の変化や、エッジが少ない、存在しないアイテムに対してもロバストになる。
Decoderでは、以下の式で対象のcompatibilityを計算
スクリーンショット 2019-04-29 13 16 16

スクリーンショット 2019-04-29 13 10 29

結果

FITBとoutfit compatibilityでSOTA。(下図はpolyvoreの例)

スクリーンショット 2019-04-29 13 19 40

contextを変えると結果も変わる
スクリーンショット 2019-04-29 13 22 13

ただ、すでにcompatibilityが他アイテムと計算されている(すでにコーディネートで使われている、購入の共起があるなど)ことを前提としているため、完全に新商品同士とかだとそのままは使えないはず。

author

Guillem Cucurull
Element AI
gcucurull@elementai.com
Perouz Taslakian
Element AI
perouz@elementai.com
David Vazquez
Element AI
dvazquez@elementai.com

URL

https://arxiv.org/pdf/1902.03646.pdf

year

2019

Metadata

Metadata

Assignees

No one assigned

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions