The Role of Traditional Concepts in the Era of Learning-Based Scene Modeling With Neural Radiance Fields and Gaussian Splatting: Why Classical Concepts Still Matter

Neural Radiance Fields (NeRFs) and, more recently, 3-D Gaussian Splatting (3DGS) have emerged as powerful learning-based approaches for scene representation, reconstruction, and novel view synthesis. While often presented as new paradigms, their success relies on concepts that have shaped the field for decades, including sampling and reconstruction theory, kernel-based rendering and splatting, geometric modeling, robust estimation, uncertainty modeling, physically motivated appearance modeling, and semantic constraints. We argue that learning-based scene modeling has not rendered these classical ideas obsolete but rather benefits from incorporating them as inductive biases. Using recent NeRF- and Gaussian-Splatting-based methods as examples, we show that many advances arise from rediscovering, reformulating, or encoding traditional concepts within modern, differentiable optimization frameworks. This perspective suggests that incorporating geometric, statistical, semantic, and physical constraints remains essential for improving robustness and generalization in learning-based scene representations, particularly in challenging settings.