【ECCV 2018】Implicit 3D Orientation Learning for 6D Object Detection from RGB Images (Oral; Best Paper)
1. Implicit 3D Orientation Learning for
6D Object Detection from RGB Images
(ECCV 2018 Oral; Best Paper)
Martin Sundermeyer1, Zoltan-Csaba Marton1, Maximilian Durner1,
Manuel Brucker1, Rudolph Triebel1,2
1German Aerospace Center (DLR), 2 Technical University of Munich
1
http://hirokatsukataoka.net/project/cc/index_cvpaperchallenge.html
資料作成:⽚岡 裕雄
6. 関連研究(3/4)
6
• Domain Randomization
– CAD学習(下図),実空間6D検出(右図)
– ベースモデル: Faster R-CNN
– 学習パラメータ固定*発⾒
Hinterstoisser+, “On Pre-Trained Image Features and Synthetic Images for Deep Learning,” in arXiv
preprint 1710.10710, 2017. https://arxiv.org/pdf/1710.10710.pdf
* CADの⼤量データで学習したパラメータを固定,出⼒に近い
CNNパラメータのみをFinetuning
7. 関連研究(4/4)
7
• SSD-6D
– 物体検出 => 3D 姿勢検出という流れ
– 異なる⼤きさの物体をSSDでマルチスケール検出
– 3D Rotationの学習をCADで実施
Kehl+, “SSD-6D: Making RGB-Based 3D Detection and 6D Pose Estimation Great Again,” in ICCV, 2017.
http://openaccess.thecvf.com/content_ICCV_2017/papers/Kehl_SSD-6D_Making_RGB-
Based_ICCV_2017_paper.pdf
15. Ablation Study
15
• Latent Space Size
• CAD vs. Textured 3D Reconstruction
AAEの潜在空間のサイズ
• 64-dimで頭打ち
モデルはCADか?3Dモデルの2D投影?
• 学習回数25,000までを評価
• Textured 3D Reconstructionの勝利