Efficient 3D Object Synthesis and Modeling Through Generative Adversarial Networks

The synthesis and modeling of 3D object is many a time a crucial need in vision and graphics applications. Traditionally, generating 3D representations involves the use of infrared or depth-sensing cameras to establish a foundational structure, followed by manual refinement using 3D modeling software such as CAD and Blender. This method is labor-intensive. Therefore, exploring more efficient 3D object synthesis and modeling methods is imperative in the realms of graphics and computer vision. Employing learning-based techniques particularly generative adversarial networks (GANs), offers a promising avenue. GAN undergoes training in an adversarial setup across multiple iterations which enable it to learn features from the training dataset to a degree where distinguishing between dataset points and a generated output becomes challenging. Leveraging this capability of GANs facilitates the optimization of the object modeling process, allowing for the unsupervised creation of object models in three-dimensional space with minimal or no human intervention. In this manuscript, we introduce a Generative Adversarial Network (GAN) model designed for the generation of 3D images from 2D inputs. Our approach involves creating a lightweighted generator and a discriminator model to generate an efficient 3D object using 2D object with minimum time. The generator incorporates a 3D shape generator, viewpoint generator, and projection module to produce a three-dimensional representation of an object. This paper has also covered a comparative analysis of proposed model with the previous model by evaluating and comparing chamfer distance of various object.