WebSVR: Joint Volume Measurement and Novel View Synthesis in Web-based 3D Environments

Abstract

Traditional 3D representation models, such as 3D meshes and 3D point clouds, provide accurate geometric measurements but lack the visual fidelity of modern novel view synthesis approaches. In contrast, recent differentiable rendering methods, such as Plenoxels and Gaussian Splatting, enable photorealistic real-time visualization but do not support common geometric measurement operations. This work builds on Sparse Voxel Rasterization (SVR), a differentiable rendering method that combines the visual fidelity of Gaussian Splatting with an adaptive voxel representation. Our first contribution is VolSVR, a volume-preserving modification of SVR training that retains voxels inside objects while maintaining accurate scene representations. Our second contribution is WebSVR, a browser-based SVR viewer for interactive visualization and volume measurement. Together, these contributions form a semi-automated workflow for volume measurement in real-time rendered scenes with high visual fidelity. The proposed method requires only 100–200 input images, does not increase SVR runtime, and achieves volume estimates with an approximately \(17 \,\%\) mean absolute error. The approach is robust to object occlusion and is scale-agnostic, supporting measurements ranging from hundreds of milliliters to thousands of cubic meters.