A Decade of Deep Learning-Based 2D-to-3D Image Reconstruction: Taxonomy, Challenges, and Emerging Paradigms

In recent years, two-dimensional to three-dimensional (2D-to-3D) image reconstruction has become one of the main research problems in the field of computer vision due to the development of deep learning. The survey will overview the development of key learning paradigms in the reconstruction of 2D-to-3D (convolutional neural networks (CNNs), generative adversarial networks (GANs), support vector machines (SVMs), and how they are used in hybrid reconstruction models). The methods reviewed are discussed concerning the ability to extract features, the ability to perceive and the ability to compute. Special focus is given to hybrid systems which comprise CNN-based learning of spatial representations and GAN-based learning of generative improvement and SVM-based learning of discriminative decision support to enhance fidelity to replication. Such important issues as depth ambiguity, scalability, and cross dataset generalization are discussed systematically. The survey also identifies the emerging paradigms, such as Neural Radiance Fields (NeRFs), Vision Transformers (ViTs), and diffusion-based models that have a high potential to represent the implicit aspects of scenes and generate high-fidelity 3D scenes. In general, this review seeks to offer a systematic reference to any future research on interpretable, scalable and robust 3D vision systems.