A Novel Pipeline for 3D Gaussian Splatting: Bridging Path Tracing and Real-Time Radiance Fields

Novel View Synthesis techniques such as 3D Gaussian Splatting (3DGS) conventionally rely on Structure-from-Motion (SfM) to generate the sparse point cloud required for initialisation. SfM, however, fails systematically on textureless or flat surfaces, producing sparse geometry and visible artefacts. This paper presents a synthetic data generation pipeline that bypasses SfM entirely by decoupling geometric and photometric acquisition. A custom path tracing engine employs a toroidal sensor to capture omnidirectional data and produces a dense, ground-truth point cloud that serves as an optimal initialisation state for 3DGS. We investigate several surface sampling strategies and demonstrate that Colour-Based Importance Sampling outperforms uniform methods by concentrating samples on visually informative regions. Experimental results suggest that, within our synthetic setup, our pipeline can operate without Adaptive Density Control, achieves competitive reconstruction quality compared to standard SfM-based initialisations on the evaluated scenes, and matches COLMAP training times despite operating on a significantly denser primitive set.