Twin4CH

Urban heritage increasingly faces threats from natural disasters and anthropogenic activities, demanding preservation techniques that balance affordability, speed, and accuracy. Real-time rendering with 3D Gaussian Splatting (3DGS) offers immense potential for rapid heritage documentation; however optimal training configurations especially for aerial nadiral images remain limited. This study compares three 3DGS implementations: native 3DGS, Postshot Splat3, and MCMC variants against conventional multi-view stereo using an 80-meter altitude drone aerial imagery. Results reveal some interesting key findings where moderate downsampling of 1/8 resolution can substantially outperform full resolution training. Postshot Splat3 achieved optimal photometric quality while training 21x faster than native implementations. However, geometric analysis showed that rendering quality and geometric accuracy are not correlated. Furthermore, the result suggested that MCMC configurations failed to compete with native 3DGS and Postshot Splat3. These findings suggest that 3DGS-based heritage documentation requires a rather conservative resolution selection and optimized implementation for efficiency to allow cost effective preservation while maintaining dimensional fidelity.

The use of digital techniques has seen an increasing amount of use in recent years for heritage documentation. The development of artificial intelligence (AI) also contributed to this rise, with many different applications to help facilitate the heritage recording process. A by-product of these developments is the increasing amount of available data, in tandem with the ever-increasing need for training data for AI purposes. This paper aims to re-use old datasets and repurpose them using modern methods. The objective is therefore to see if older datasets may be used to improve the quality of AI-based methods, while also investigating the use of new technologies such as Visual Language Models (VLM) to perform semantic queries and Gaussian splatting on these datasets. For this purpose, datasets from a previous documentation project involving Javanese “candi” architecture is used in this paper since this particular subject has not seen too many AI-based documentation research in the literature and is thus an interesting example to evaluate the generalisation of AI methods. Results show that old datasets can very well be used with modern techniques with promising results. In terms of semantic segmentation, machine learning yielded an overall accuracy of 0.89 while deep learning yielded 0.79. Several interesting inferences were also observed in the VLM query results, while Gaussian splatting showed very strong potential for visualisation-based applications to further enhance the reusability of these old datasets.

The notion of radiance fields is a very recent topic which has gained much traction in the past couple of years. Among some of the well known techniques, 3D Gaussian splatting (3DGS) seems to generate promising results in terms of 3D reconstruction. In the domain of heritage documentation, orthophotos are a standard product usually generated from photogrammetry by projecting image pixels into a 3D surface created from depth maps or point clouds. In this paper we investigate the possibility to substitute the conventional method for 3D surface generation, either terrestrial laser scanning (TLS) or multi-view stereo (MVS), with 3DGS. The paper will attempt to answer two questions: (1) whether this novel method is a viable solution for orthophoto generation within the heritage documentation context, and (2) at what point is 3DGS worth the time and effort to reach an acceptable orthophoto when compared to conventional methods. To do this, we perform multiple analysis on both the geometric quality of point clouds and the orthophotos themselves, using conventional TLS data as our reference. Our results indicate that 3DGS is not only promising, but indeed a viable alternative to conventional MVS with a potential gain of processing time between 12% up to 33% to reach a comparable quality. In the scenarios tested, both of these gains came with an average point cloud standard deviation of 2 cm.