3D Object Reconstruction from 2D Images Using Variational Autoencoders (VAE)
Aezeden O. Mohamed, Rainier Nii, Kipas Binga, Alok Kumar Pandey, J. Karpagam, T J Nandhini · 2025
In this research, we present a new method on how to reconstruct 3D object from 2D images by employing Variational Autoencoders (VAE). His generality is a bit stark; however, it is well justified given that classical approaches to 3D reconstruction from multiple perspectives fail to capture detailed geometries from few 2D views. These limitations are surmounted by the proposed VAE based framework our approach learns the distribution of 3D shapes using probabilistic inference leading to accurate reconstruction from one or two 2D images. The model also has a decoder that provides a 3D mesh or voxel grid to effectively be used in computer vision, AR and VR applications. The VAE model takes benefit of a big dataset of 3D object shapes, while the encoder dimension translate the 2D picture in the latent space and the decoder rebuild the 3D objects from this space. From the proposed method evaluation, the results suggest high reconstruction accuracy of the proposed method as compared to the traditional point cloud and mesh-based methods in both quality and time taken. The VAE-based system also improves the model's generalization because it has low computational complexity and creates clear and realistic 3D models. The suggested approach has a massive application to the real-time 3D object reconstruction in the different areas including games, product designing as well as diagnosing tools in medical sciences.