Title: Recent Developments and Trajectories in Learning-Driven 3D Representations

Abstract: Choosing the right 3D representation is a critical architectural choice that determines the performance, fidelity, and functional scope of contemporary computer vision and graphics systems. These systems support essential operations including 3D reconstruction, novel-view synthesis, rendering, shape and motion analysis, recognition, and generation. Although conventional formats such as point clouds, meshes, and volumetric grids continue to serve as the standard outputs for 3D sensing hardware like LiDAR and 3D scanners—and remain prevalent in downstream tasks such as simulation and editing—emerging neural and primitive-based approaches, including 3D Gaussian Splatting, provide efficient and differentiable alternatives. These newer methods unlock significant potential across diverse sectors, ranging from autonomous driving and AR/VR to medical imaging, robotics, and gaming. This study surveys the primary categories of 3D representations, spanning from discrete, explicit structures to continuous implicit fields grounded in neural rendering or primitive splatting. For every representation class, the paper details its core formulation and variations, evaluates its advantages and drawbacks, and identifies its principal use cases. The discussion concludes by addressing unresolved issues and suggesting avenues for future inquiry. Unlike previous reviews that offer a broad overview of 3D object and scene reconstruction, this work delivers a concentrated examination of the progression of the representations themselves. It underscores the transition toward implicit models, presenting a fresh viewpoint on how these innovative formats are reshaping 3D and 4D workflows at a fundamental level.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC