| Title |
Improved ResNet VAE-GAN 3D Object Creation System using Residual Blocks |
| Authors |
박세영(Seayoung Park) ; 안성수(Sungsoo Ahn) ; 이정석(Jungsuk Lee) ; 윤대열(Daiyeol Yun) |
| DOI |
https://doi.org/10.5573/ieie.2024.61.10.79 |
| Keywords |
Voxel; Mesh; 3D object generation; VAE; GAN; Residual block |
| Abstract |
3D modeling is used in various fields such as games, AR, VR, and metaverse. Recently, visualization and computation in 3D space are accelerating due to improvements in the performance of computer hardware, and methods for generating 3D objects are being researched due to advances in GAN technology. The GAN-based network proposed a 3D-VAE-IWGAN method that receives images as input, generates voxels, and learns by introducing a Wasserstein loss function and applying a gradient penalty. GAN can generate multiple models that are not included in training, but there is a problem with artifacts occurring. As another method, a network such as DIB-R, which reduces the cost of 3D label generation through supervised learning in 2D and unsupervised learning in 3D, has been proposed. DIB-R can reduce artifacts, but it is difficult to generate diverse models with an autoencoder-based network. This paper proposes a method of applying residual blocks to VAE-GAN, which combines Variational Autoencoder (VAE) and Generative Adversarial Network (GAN), which improves performance in 3D-VAE-IWGAN, and further improves the image generator and discriminator. We propose a system that extracts many features to generate high-quality images and improve latent space interpolation performance. The results of index compared to the existing network showed a better result of 116.33% at 137.15 in the chair class, and an improvement of 130.4% at 137.24 in the bed class. |