Conference Agenda

3D Gaussian Splatting (3DGS) has revolutionized real-time rendering with state-of-the-art novel view synthesis, but its applicability to accurate geometric measurement remains limited. Compared with multi-view stereo (MVS)-based point clouds or mesh models, 3DGS provides superior visual quality and completeness, while existing measurement approaches still rely on stereoscopic workstations or direct measurements on incomplete and inaccurate reconstructed geometry.

As a novel view synthesizer, 3DGS reproduces source views and smoothly interpolates intermediate viewpoints, enabling users to intuitively identify congruent points across multiple views. By triangulating these correspondences, accurate 3D point measurements can be obtained. Inspired by traditional stereoscopic measurement, the proposed approach removes the need for stereo workstations and biological stereoscopic capability, while naturally supporting multi-view measurements for improved accuracy.

We implement a web-based application to demonstrate this proof of concept using UAV-based aerial datasets. Experimental results show that the proposed method achieves measurement accuracy comparable to or better than traditional stereoscopic measurement approaches while operating entirely on non-stereo workstations. In particular, the proposed method consistently outperforms direct mesh-based measurements, achieving RMSEs of 1–2 cm on well-defined points. On challenging thin structures, the proposed method reduces RMSE from 0.062 m to 0.037 m, and successfully measures sharp corners where mesh-based methods fail entirely.

The source code and documentation are open-source and available at: https://github.com/GDAOSU/3dgs_measurement_tool.

Traditional 3D Gaussian Splatting (3DGS) methods initialize Gaussian primitives from Structure-from-Motion point clouds, resulting in loosely distributed representations that lack geometric constraints and metric accuracy. This limitation severely restricts their application in architectural heritage preservation, where millimeter-level precision and practical editability are essential requirements.

This paper introduces Gaussian Texturing, a novel framework that fundamentally transforms how Gaussians relate to geometry by directly binding 3D Gaussian primitives to precisely measured mesh surfaces—essentially "texturing" surfaces with Gaussians. Our approach comprises three key innovations: (1) a constrained optimization framework that maintains tight Gaussian-surface coupling throughout training, preventing geometric drift while preserving photorealistic rendering quality; (2) engineering-oriented editing tools enabling geometry-based material replacement, region editing, and mesh-driven deformation; and (3) seamless integration with professional heritage preservation workflows.

Experimental validation on MipNeRF360 benchmarks and custom architectural datasets demonstrates that our method achieves millimeter-level geometric precision while maintaining competitive rendering metrics. Unlike traditional "bind-after-training" approaches, our direct surface binding paradigm eliminates intermediate reconstruction steps, ensuring accuracy from source data. Real-world applications in heritage documentation and architectural design confirm the method's practical value, successfully bridging the gap between photorealistic visualization and engineering-grade geometric accuracy for professional applications.

In this study, we develop a Structured framework for Gaussian Splatting (3DGS) with LiDAR integration (Structured-Li-GS). It is a lightweight Gaussian Splatting pipeline that leverages LiDAR–inertial–visual SLAM. Structured-Li-GS achieves high-quality 3D reconstructions with fewer Gaussians by training on accurate, dense, colorized point clouds. Gaussian primitives are anchored using sub-sampled point clouds, and their ellipsoidal parameters are initialized from local surface geometry. Our training strategy integrates a comprehensive set of loss terms, including photometric, flattening, offset, depth, and normal losses—guided by the dense point cloud, enabling accurate reconstruction without Gaussian densification. This approach produces up-to-scale, high-fidelity results with a moderate model size. For experimental validation, we develop a custom hardware-synchronized LiDAR–camera handheld scanner. Experiments on both benchmark datasets and our real-world in-house dataset demonstrate that Structured-Li-GS surpasses state-of-the-art methods while using fewer Gaussians.

True Digital Orthophoto Maps (TDOMs) are essential for urban analysis and map updating, traditionally generated through photogrammetric workflows involving aerial triangulation, DSM construction, and orthorectification. Recently, 3D Gaussian Splatting (3DGS) has emerged as an alternative approach using differentiable volumetric rendering. While both methods depend on acquisition geometry, they follow fundamentally different reconstruction processes, potentially producing distinct representational characteristics. Systematic comparisons under controlled conditions remain limited.

This study generates photogrammetric and 3DGS-based TDOMs from four UAV datasets acquired over the same area with varying resolution (2.51–5.8 cm GSD), image count, and oblique view proportion (0–75%). All datasets were preprocessed through common SfM to obtain identical inputs. We evaluate differences through inter-method agreement (PSNR, SSIM, LPIPS), detail preservation (gradient magnitude, high-frequency energy), and spatial distribution patterns (boundary–interior separation).

Results show 3DGS systematically smooths fine-scale texture with gradient ratios of 0.58–0.89 and high-frequency energy reductions of 2.5–55× relative to photogrammetry. Oblique view proportion emerges as the dominant divergence factor: oblique-dominant datasets show lowest agreement (PSNR 15.15) despite larger image counts, while nadir-only datasets achieve higher similarity (PSNR 26.73). Difference maps reveal 2–3 times higher discrepancies along boundaries than interiors. Visually cleaner 3DGS boundaries are byproducts of overall smoothing rather than superior reconstruction. These findings establish that the two methods are complementary—photogrammetry preserving texture fidelity and 3DGS providing structural regularity—with acquisition geometry critically influencing performance characteristics.

Efficient and robust 3D scene representation is crucial in fields such as robotics, autonomous driving, and augmented reality. While RGBimagesprovidevaluable content for 3D reconstruction, other modalities like thermal or depth can enable additional information on the 3D environment. Lately, Novel View Synthesis (NVS) methods like Gaussian Splatting (GS) have started using multiple modalities to further boost their performance. But fusing or combining those multi-modal data can make the process slower and bring in additional challenges. Therefore, our project aims to use single modality based on thermal infrared domain, by removing the reliance on visible light, as much as possible. We propose a method Thermal-to-Depth Gaussian (TDg), that uses only thermal images and depth estimation in its architecture to derive the radiance fields. Mainstream methods relying heavily on RGB images, perform poorly in visually degraded environments, such as low-light conditions, fog, smoke, or extreme weather. Contrary to this, infrared cameras can detect objects’ inherent thermal radiation and provide a robust perception, suitable regardless of lighting and weather conditions. But despite their promise, thermal images are inherently characterized by low contrast, sparse texture, and non-uniform brightness distribution. So current approaches still rely heavily on paired RGB images for supervision or joint optimization, failing to establish a truly independent and purely thermal-based Gaussian representation system. Therefore, the core innovation of our work is to prepare a self contained Thermal GS framework that uses only thermal image inputs. We design a thermal-guided depth estimation module, Thermal-to-Depth (TDg), providing explicit and reliable constraints for geometric optimization.