Papers of the day   All papers

PolyGen: An Autoregressive Generative Model of 3D Meshes

Comments

Charlie Nash: Excited to share PolyGen, a generative model of 3D meshes. PolyGen sequentially generate meshes vertex-by-vertex, and face-by-face using Transformers. With @yaroslav_ganin @arkitus @PeterWBattaglia @DeepMind http://arxiv.org/abs/2002.10880 https://t.co/uBUUxPssth

13 replies, 858 likes


DeepMind: PolyGen: an autoregressive generative model of 3D meshes. PolyGen sequentially generates meshes vertex-by-vertex and face-by-face, using Transformers and Pointer Networks. http://arxiv.org/abs/2002.10880 @charlietcnash @yaroslav_ganin @arkitus @PeterWBattaglia https://t.co/qWLv9A3Fdj

2 replies, 401 likes


Ali Eslami: Introducing PolyGen: an autoregressive model of 3D meshes. https://arxiv.org/abs/2002.10880 Transformers + Pointer Nets = train on raw mesh data (i.e. variable-length lists of vertices and faces). No need to voxelise or rasterise! with @charlietcnash @yaroslav_ganin @PeterWBattaglia https://t.co/oaSluTXEay

5 replies, 308 likes


Eric Jang πŸ‡ΊπŸ‡ΈπŸ‡ΉπŸ‡Ό: Wow! I used to think that we needed RL agents to learn structured synthesis with non-differentiable objectives/constraints (e.g. obeying mesh closure). but maybe all you really need is max likelihood and a clever factorization of the problem.

3 replies, 126 likes


Oriol Vinyals: Amazing how far mesh generation has been pushed! I post a video from Pointer Networks (2015), in which results like the one below by @charlietcnash et al seemed impossible. https://t.co/pPkwczl1NO

0 replies, 99 likes


Yaroslav Ganin: PolyGen: a generative model for 3D meshes β€’ Outputs vertices one by one then generates faces (all using Transformers) β€’ Faces are polygons; more structure πŸ‘‰ better results β€’ Pointer nets to select vertices for each face Led by @charlietcnash. w/ @arkitus @PeterWBattaglia

1 replies, 64 likes


Meire Fortunato: Impressive results by @charlietcnash et al. Mesh generation by solving numerical PDEs (no ML) was my PhD thesis topic. Pointer Networks(2015) was my first paper in ML, there we only attempted generating very simple 2D triangulations in a seq2seq fashion - progress is exciting! :)

0 replies, 39 likes


Andrew Davison: Generative models of meshes look promising for modelling complicated objects and I'll read this one properly.

1 replies, 32 likes


Xander Steenbrugge: Some amazing new 3D-modelling tools incoming πŸ‘Œ #Blender3d

0 replies, 26 likes


Daisuke Okanohara: PolyGen is a generative model of 3D polygon meshes. It uses an autoregressive model with a transformer and pointer network, which generates vertices and faces sequentially. Generation can condition on class and images. https://arxiv.org/abs/2002.10880

0 replies, 7 likes


AI Core: Humans are constrained to thinking in simple shapes when designing hence everyday objects take very simple, non optimal shapes. PolyGen+Reinforcement Learning combined with 3d printing is going to revolutionize the design of these forms.

0 replies, 2 likes


Toni Rosinol: I’m just amazed at what has been achieved by @charlietcnash et al., seems like we are getting closer to having a 3D mesh reconstruction of the world, would be interesting to see how these principles apply to SLAM though.

0 replies, 2 likes


Matthew Vowels: Using graph NNs effectively for CAD furniture design design https://arxiv.org/pdf/2002.10880.pdf

0 replies, 1 likes


Content

Found on Feb 26 2020 at https://arxiv.org/pdf/2002.10880.pdf

PDF content of a computer science paper: PolyGen: An Autoregressive Generative Model of 3D Meshes