UCSG-Net - Unsupervised Discovering of Constructive Solid ...
1
UCSG-N ET - Unsupervised Discovering of Constructive Solid Geometry Tree Kacper Kania 1 , Maciej Zieba 1,2 , Tomasz Kajdanowicz 1 1 Wroclaw University of Science and Technology 2 Tooploox [email protected] Contributions I The first method that learns CSG operations in unsupervised manner. I Predicted CSG trees are fully interpretable and controllable. They can aid design of 3D objects. I In terms of reconstruction quality, our model is on par with existing methods that aim for interpretable 3D object reconstruction. Motivation CSG is a common tool that can be in 3D graphics software for modeling objects with complex topology. We aim to automate the process by pre- dicting CSG trees that create these ob- jects. ∪ * ∩ * - * Existing approaches Time consuming inference InverseCSG, Du et al., SIGGRAPH 2018 Require supervision of CSG operations CSG-Net, Sharma et a., CVPR 2018 Learnable CSG Layer Masking ˆ V Relations Output shapes from layer l - 1 Initial shapes Output shapes from layer l A ∪ * BA ∩ * B A - * BB - * AA ∪ * BA ∩ * BA - * BB - * A A ∪ * BA ∩ * BA - * BB - * A I Masks (separate for left and right operands of CSG operations) select pair of elements: V left = softmax(K left z) V right = softmax(K right z) I Masked shapes ˆ V side,i are obtained by sampling Gumbel-Softmax. I Operands for l -th layer are obtained as: A = O left = M X i =1 O i ˆ V left,i B = O right = M X i =1 O i ˆ V right,i Representing a single shape and CSG operations CSG operations for occupancies A ∪ * B :[ + ] [0,1] = A ∩ * B :[ + - 1] [0,1] = A - * B :[ - ] [0,1] = B - * A :[ - ] [0,1] = UCSG-N ET Training UCSG-N ET First stage Second stage - when α ≤ 0.05 Results - 2D CAD Dataset Method Mode k i = 0 i = ∞ CSG-N ET S TACK Supervised 1 3.98 2.25 CSG-N ET S TACK Supervised 10 1.38 0.39 CSG-N ET S TACK RL 1 1.27 0.57 CSG-N ET S TACK RL 10 1.02 0.34 Our Unsupervised 1 0.32 - ∪ * Ground truth Reconstruction Primitives CSG Tree - * ∪ * ∪ * ∪ * - * - * ∪ * ∪ * ∪ * ∪ * - * ∪ * - * ∪ * Results - 3D ShapeNet CD - Chamfer Distance ×10 3 High interpretability Low interpretability Ours VP SQ BAE BSP-Net CD 2.085 2.259 1.656 1.592 0.446 ∩ * ∩ * Primitives ∩ * ∩ * ∪ * ∪ * ∪ * ∩ * ∩ * ∩ * ∩ * - * ∪ * ∩ * ∩ * Primitives ∩ * ∩ * ∪ * ∪ * ∪ * ∩ * ∩ * ∩ * ∩ * - * ∪ * ∩ * ∩ * Primitives ∩ * ∩ * ∪ * ∪ * ∪ * ∩ * ∩ * ∩ * ∩ * - * ∪ * Difference Intersection Union ∩ * ∩ * Primitives Layer 1 Layer 2 Layer 3 Layer 5 Layer 4 ∩ * ∩ * ∪ * ∪ * ∪ * ∩ * ∩ * ∩ * ∩ * - * ∪ * Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS 2020)
Transcript of UCSG-Net - Unsupervised Discovering of Constructive Solid ...
UCSG-NET - Unsupervised Discovering of Constructive Solid Geometry TreeKacper Kania1, Maciej Zieba1,2, Tomasz Kajdanowicz1
1Wroclaw University of Science and Technology [email protected]
Contributions
I The first method that learns CSG operations in unsupervised manner.I Predicted CSG trees are fully interpretable and controllable. They can aid design
of 3D objects.
I In terms of reconstruction quality, our model is on par with existing methods thataim for interpretable 3D object reconstruction.
Motivation
CSG is a common tool that can be in 3Dgraphics software for modeling objectswith complex topology.
We aim to automate the process by pre-dicting CSG trees that create these ob-jects.
∪∗
∩∗
−∗
Existing approaches
Time consuming inference
InverseCSG, Du et al., SIGGRAPH 2018
Require supervision of CSG operations
CSG-Net, Sharma et a., CVPR 2018
Learnable CSG Layer
Masking V̂
Relations
Output shapes from layer l − 1Initial shapes
Output shapes from layer l
A ∪∗ B A ∩∗ B A−∗ B B −∗ A A ∪∗ B A ∩∗ B A−∗ B B −∗ A A ∪∗ B A ∩∗ B A−∗ B B −∗ A
I Masks (separate for left and right operands of CSG operations) select pair ofelements:
Vleft = softmax(Kleftz) Vright = softmax(Krightz)I Masked shapes V̂side,i are obtained by sampling Gumbel-Softmax.
I Operands for l-th layer are obtained as:
A = Oleft =M∑
i=1
OiV̂left,i B = Oright =M∑
i=1
OiV̂right,i
Representing a single shape and CSG operations
CSG operations for occupancies
A ∪∗ B : [ + ][0,1] =
A ∩∗ B : [ + − 1][0,1] =A−∗ B : [ − ][0,1] =
B −∗ A : [ − ][0,1] =
UCSG-NET
Training UCSG-NET
First stage
Second stage - when α ≤ 0.05
Results - 2D CAD Dataset
Method Mode k i = 0 i =∞CSG-NETSTACK Supervised 1 3.98 2.25CSG-NETSTACK Supervised 10 1.38 0.39CSG-NETSTACK RL 1 1.27 0.57CSG-NETSTACK RL 10 1.02 0.34
Our Unsupervised 1 0.32 -
∪∗
Ground truth
Reconstruction
Primitives
CSG Tree
−∗ ∪∗
∪∗ ∪∗
−∗ −∗
∪∗
∪∗
∪∗
∪∗−∗
∪∗
−∗ ∪∗
Results - 3D ShapeNet
CD - Chamfer Distance ×103
High interpretability Low interpretabilityOurs VP SQ BAE BSP-Net
CD 2.085 2.259 1.656 1.592 0.446
∩∗
∩∗
Pri
mit
ives
∩∗∩∗
∪∗
∪∗
∪∗∩∗
∩∗
∩∗∩∗
−∗
∪∗
∩∗
∩∗
Pri
mit
ives
∩∗∩∗
∪∗
∪∗
∪∗∩∗
∩∗
∩∗∩∗
−∗
∪∗
∩∗
∩∗
Pri
mit
ives
Layer 1 Layer 2 Layer 3 Layer 5Layer 4
∩∗∩∗
∪∗
∪∗
∪∗∩∗
∩∗
∩∗∩∗
−∗
∪∗
DifferenceIntersectionUnion
∩∗
∩∗
Primitives
Layer 1 Layer 2 Layer 3 Layer 5Layer 4
∩∗∩∗
∪∗
∪∗
∪∗∩∗
∩∗
∩∗∩∗
−∗
∪∗
Thirty-fourth Conference on Neural Information Processing Systems (NeurIPS 2020)
