Convergence behaviour and computational efficiency of Gaussian Splatting in complex heritage environments

This study evaluates the performance of three-dimensional Gaussian Splatting (3DGS) for the documentation and virtualisation of architectural cultural heritage. Six datasets acquired in different areas of the Mosque–Cathedral of Córdoba were analysed, representing a variety of capture strategies and scene scales. The datasets were obtained using different acquisition devices, including a drone, a Digital Single-Lens Reflex (DSLR) camera, a mobile phone, and a 360° camera. The images were photogrammetrically orientated using a Structure-from-Motion (SfM) workflow in RealityScan, and the estimated camera intrinsics and poses were subsequently exported to COLMAP format. The datasets were processed using 3DGS, and the training behaviour was analysed across two hardware configurations (RTX3070Ti and RTX5070 GPUs). The study evaluated reconstruction quality and convergence behaviour using three complementary image quality metrics: Peak Signal-to-Noise-Ratio (PSNR), Structural Similarity Index Measure (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS). Additionally, training time and computational efficiency were analysed to determine how scene complexity and hardware configuration affect the optimisation process. The results show that most of the improvement in perceptual quality occurs during the early stages of training. In the majority of cases, 90–95% of the perceptual improvement is achieved between 10000 and 20000 iterations, while the remaining iterations contribute mainly to minor visual refinements. The analysis also demonstrates that the hardware configuration does not significantly influence the final reconstruction quality, although it strongly affects computational performance in large scenes due to the progressive densification of Gaussian primitives during training. Overall, the results suggest that the number of training iterations can be significantly reduced without substantial loss of perceptual quality. These findings highlight the potential of GS as an efficient methodology for the digital documentation and visualisation of architectural heritage, combining high visual fidelity, relatively short training times, and the capacity for interactive rendering for research, dissemination, and educational applications.

Convergence (economics) · Gaussian · Image (mathematics) · Intrinsics · Perception · Similarity (geometry) · Workflow · 3D Surveying and Cultural Heritage · Advanced Image and Video Retrieval Techniques · Advanced Vision and Imaging

• Image quality assessment

  Open Access•Zhou Wang, Zhou Wang Zhou Wang et al.•IEEE Transactions on Image…•2004

  Add to list

• D reconstruction of cultural heritage sites; A case study of Patan Durbar Square

  Open Access•Pratik Shrestha, Subarna Prasad Kapali et al.•Digital Applications in…•2025

  Add to list