Abstract
Lowering the costs of generating realistic indoor scenes upon customer’s requests, automatic synthesis of diverse indoor furniture layouts opens the door to various innovative business models for the real estate and home improvement industry. With the continuous development of deep generative networks, the quality of their generation results is constantly improving and occupies an important position in industrial design. Therefore, how to utilize deep generative networks to design indoor 3D scene layout with high quality and high level is of great significance to improve the design efficiency. This paper introduce RoomGen, a 3D Generative Adversarial Network (GAN) tailored for this purpose. RoomGen integrates an enhanced Variational Autoencoder (VAE) and a UNet++ discriminator for discriminate feature learning. Specifically, VAE introduces residual module in the encoder and deep supervision in the decoder, which improves the fidelity of the generated layouts. In addition, RoomGen incorporates a novel attention mechanism in the UNet++ recognizer, which enables it to capture fine-grained features and better distinguish between real and generated layouts. Experimental results demonstrate RoomGen can proficiently generate high-quality indoor furniture layouts with impressive reconstruction accuracy. Additionally, RoomGen’s capability to produce diverse and realistic scenes offers a more cost-effective solution for the real estate and home improvement sectors. Moreover, it provides a fully automated solution for designing virtual reality scenes and 3D game environments.