Open Yagami360 opened 4 years ago
従来の DenseReg での手法(Fully-convolutional dense pose regression) 従来の DenseReg では、2D画像の UV 座標から3D形状の表面へのテスクチャマッピングを、分類+回帰を行う全畳み込みネットワーク(FCN)でモデル化している。 即ち、まず分類タスクでは、2D画像の該当 UV 座標が、3Dの背景に属するのか3D形状の特定の人体部位表面へ属するのかを分類する。 次に、回帰タスクでは、2D画像の該当 UV 座標から3D形状の表面座標への回帰を行う。
このことを定式化すると、以下の式のようになる。
Region-based Dense Pose Regression
Multi-task cascaded architectures
Distillation-based ground-truth interpolation
0. 論文情報・リンク
1. どんなもの?
2. 先行研究と比べてどこがすごいの?
3. 技術や手法の"キモ"はどこにある?
従来の DenseReg での手法(Fully-convolutional dense pose regression) 従来の DenseReg では、2D画像の UV 座標から3D形状の表面へのテスクチャマッピングを、分類+回帰を行う全畳み込みネットワーク(FCN)でモデル化している。 即ち、まず分類タスクでは、2D画像の該当 UV 座標が、3Dの背景に属するのか3D形状の特定の人体部位表面へ属するのかを分類する。 次に、回帰タスクでは、2D画像の該当 UV 座標から3D形状の表面座標への回帰を行う。
このことを定式化すると、以下の式のようになる。
Region-based Dense Pose Regression
Multi-task cascaded architectures
Distillation-based ground-truth interpolation
4. どうやって有効だと検証した?
5. 議論はあるか?
6. 次に読むべき論文はあるか?
7. 参考文献