Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Standard

Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds. / Saha, Sneha; Menzel, Stefan; Minku, Leandro L.; Yao, Xin; Sendhoff, Bernhard; Wollstadt, Patricia.

IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings. IEEE Computer Society Press, 2020.

Research output: Chapter in Book/Report/Conference proceedingConference contribution

Harvard

Saha, S, Menzel, S, Minku, LL, Yao, X, Sendhoff, B & Wollstadt, P 2020, Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds. in IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings. IEEE Computer Society Press, 2020 IEEE Symposium Series on Computational Intelligence (SSCI 2020) , Canberra, Australia, 1/12/20.

APA

Saha, S., Menzel, S., Minku, L. L., Yao, X., Sendhoff, B., & Wollstadt, P. (Accepted/In press). Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds. In IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings IEEE Computer Society Press.

Vancouver

Saha S, Menzel S, Minku LL, Yao X, Sendhoff B, Wollstadt P. Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds. In IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings. IEEE Computer Society Press. 2020

Author

Saha, Sneha ; Menzel, Stefan ; Minku, Leandro L. ; Yao, Xin ; Sendhoff, Bernhard ; Wollstadt, Patricia. / Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds. IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings. IEEE Computer Society Press, 2020.

Bibtex

@inproceedings{95311847c9f940ceb2f1df438a65e38b,
title = "Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds",
abstract = "During each cycle of automotive development, largeamounts of geometric data are generated as results of designstudies and simulation tasks. Discovering hidden knowledge fromthis data and making it available to the development teamstrengthens the design process by utilizing historic informationwhen creating novel products. To this end, we propose touse powerful geometric deep learning models that learn lowdimensionalrepresentation of the design data in an unsupervisedfashion. Trained models allow to efficiently explore the designspace, as well as to generate novel designs. One popular classof generative models are variational autoencoders, which havehowever been rarely applied to geometric data. Hence, we usea variational autoencoder for 3D point clouds (PC-VAE) andexplore the model{\textquoteright}s generative capabilities with a focus on thegeneration of realistic yet novel 3D shapes. We apply the PC-VAEto point clouds sampled from car shapes from a benchmarkdata set and employ quantitative measures to show that ourPC-VAE generates realistic car shapes, wile returning a richervariety of unseen shapes compared to a baseline autoencoder.Finally, we demonstrate how the PC-VAE can be guided towardsgenerating shapes with desired target properties by optimizingthe parameters that maximize the output of a trained classifierfor said target properties. We conclude that generative modelsare a powerful tool that may aid designers in automotive productdevelopment.",
keywords = "Representation learning, geometric deep learning, point clouds, generative model, novelty",
author = "Sneha Saha and Stefan Menzel and Minku, {Leandro L.} and Xin Yao and Bernhard Sendhoff and Patricia Wollstadt",
year = "2020",
month = sep,
day = "18",
language = "English",
booktitle = "IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings",
publisher = "IEEE Computer Society Press",
note = "2020 IEEE Symposium Series on Computational Intelligence (SSCI 2020) ; Conference date: 01-12-2020 Through 04-12-2020",

}

RIS

TY - GEN

T1 - Quantifying The Generative Capabilities Of Variational Autoencoders For 3D Car Point Clouds

AU - Saha, Sneha

AU - Menzel, Stefan

AU - Minku, Leandro L.

AU - Yao, Xin

AU - Sendhoff, Bernhard

AU - Wollstadt, Patricia

PY - 2020/9/18

Y1 - 2020/9/18

N2 - During each cycle of automotive development, largeamounts of geometric data are generated as results of designstudies and simulation tasks. Discovering hidden knowledge fromthis data and making it available to the development teamstrengthens the design process by utilizing historic informationwhen creating novel products. To this end, we propose touse powerful geometric deep learning models that learn lowdimensionalrepresentation of the design data in an unsupervisedfashion. Trained models allow to efficiently explore the designspace, as well as to generate novel designs. One popular classof generative models are variational autoencoders, which havehowever been rarely applied to geometric data. Hence, we usea variational autoencoder for 3D point clouds (PC-VAE) andexplore the model’s generative capabilities with a focus on thegeneration of realistic yet novel 3D shapes. We apply the PC-VAEto point clouds sampled from car shapes from a benchmarkdata set and employ quantitative measures to show that ourPC-VAE generates realistic car shapes, wile returning a richervariety of unseen shapes compared to a baseline autoencoder.Finally, we demonstrate how the PC-VAE can be guided towardsgenerating shapes with desired target properties by optimizingthe parameters that maximize the output of a trained classifierfor said target properties. We conclude that generative modelsare a powerful tool that may aid designers in automotive productdevelopment.

AB - During each cycle of automotive development, largeamounts of geometric data are generated as results of designstudies and simulation tasks. Discovering hidden knowledge fromthis data and making it available to the development teamstrengthens the design process by utilizing historic informationwhen creating novel products. To this end, we propose touse powerful geometric deep learning models that learn lowdimensionalrepresentation of the design data in an unsupervisedfashion. Trained models allow to efficiently explore the designspace, as well as to generate novel designs. One popular classof generative models are variational autoencoders, which havehowever been rarely applied to geometric data. Hence, we usea variational autoencoder for 3D point clouds (PC-VAE) andexplore the model’s generative capabilities with a focus on thegeneration of realistic yet novel 3D shapes. We apply the PC-VAEto point clouds sampled from car shapes from a benchmarkdata set and employ quantitative measures to show that ourPC-VAE generates realistic car shapes, wile returning a richervariety of unseen shapes compared to a baseline autoencoder.Finally, we demonstrate how the PC-VAE can be guided towardsgenerating shapes with desired target properties by optimizingthe parameters that maximize the output of a trained classifierfor said target properties. We conclude that generative modelsare a powerful tool that may aid designers in automotive productdevelopment.

KW - Representation learning

KW - geometric deep learning

KW - point clouds

KW - generative model

KW - novelty

M3 - Conference contribution

BT - IEEE Symposium Series on Computational Intelligence (SSCI 2020) - Proceedings

PB - IEEE Computer Society Press

T2 - 2020 IEEE Symposium Series on Computational Intelligence (SSCI 2020)

Y2 - 1 December 2020 through 4 December 2020

ER -