[DL Hacks]Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling

3D-GAN Learning a Probabilistic Latent Space of Object Shapes via 3D Generative-Adversarial Modeling DL Hacks

• • Learning a Probabilistic Latent Space of Object Shapes via 3D GenerativeAdversarial Modeling • • Jiajun Wu, Chengkai Zhang, Tianfan Xue, William T. Freeman, Joshua B. Tenenbaum • MIT • NIPS 2016 • Arxiv URL • https://arxiv.org/abs/1610.07584

• 3D GAN • GAN • • • • 3D 3D 3D (CAD )

3D • • CAD • mesh → skeleton voxel • →3D ShapeNets • GAN

• • • • • • • • 3D convolution Conv Batch norm Batch norm ReLU Kernel 4 * 4 * 4 Strides 2 Discriminator Discriminator LRelu DCGAN

• • • •G : Adam(β=0.5) : !"#$%& = log + , + log(1 − + 1(2 ) : Generator→0.0025, Discriminator→10-5 : G 80% accuracy

3D-VAE-GAN
• VAE-GAN[Larsen et al., 2016]
•G
VAE
•
3
• GAN: !"#$%& = log + , + log 1 − + 0 1
• KL: L34 = +56 (8(1|:)||< 1 ) =

>
?(@)
∫> log A(@) B,

• Reconstruction: !CDEFG = 0 H :

−,

I

• Total: ! = !"#$%& + JK !56 + JI !CDEFG (JK = 0.5, JI = 10PQ )

• • 3D • 3D • 2D 3D

• •

• • •

• 10%

Generator • • •

Discriminator • • • object