Current browse context:
cs.CV
Change to browse by:
References & Citations
Computer Science > Computer Vision and Pattern Recognition
Title: Deep Convolutional GANs for Car Image Generation
(Submitted on 24 Jun 2020)
Abstract: In this paper, we investigate the application of deep convolutional GANs on car image generation. We improve upon the commonly used DCGAN architecture by implementing Wasserstein loss to decrease mode collapse and introducing dropout at the end of the discrimiantor to introduce stochasticity. Furthermore, we introduce convolutional layers at the end of the generator to improve expressiveness and smooth noise. All of these improvements upon the DCGAN architecture comprise our proposal of the novel BoolGAN architecture, which is able to decrease the FID from 195.922 (baseline) to 165.966.
Link back to: arXiv, form interface, contact.