🧍🏼‍♀️Build-A-Volcap🧍🏼‍♂️: Automated Synthetic Validation of Real-world Volumetric Capture Quality

Antonella Rech1, Giulia Martinelli1,2, Nicola Garau1,2, Nicola Conci1,2, Francesco De Natale1,2
1University of Trento, 2CNIT
Paper Code
Teaser image.

We present Build-A-Volcap, an automated pipeline for validating real-world Volumetric Capture setups using a synthetic Digital Twin. Left to right: scenes from a real world volumetric capture are captured and rendered to images. The original intrinsic and extrinsic parameters are extracted and transformed to build a Digital Twin of the volumetric capture. Images from the synthetic VC are then re-rendered and compared against the real ones. Exploiting this measurement, we optimize the cameras positions and validate them once again, until a desired quality in the reconstruction is reached. Finally, after the synthetic VC converges to a desirec configuration, the real VC can be adjusted accordingly.

Abstract

Volumetric capture techniques have significantly advanced in recent years, enabling detailed 3D reconstructions of dynamic real-world scenes. However, predicting their performance prior to real-world deployment remains challenging, often leading to costly experimental setups and uncertain outcomes. In this paper, we present Build-A-Volcap, an automated synthetic validation environment designed to predict real-world volumetric capture quality. Our system constructs comprehensive synthetic datasets and environments that faithfully simulate practical capture conditions. This enables rigorous testing and benchmarking of volumetric methods such as photogrammetry, NeRF, Gaussian Splatting and Radiant Foam without the need for physical setups. Through extensive experiments, we demonstrate how Build-A-Volcap effectively identifies methodological strengths and limitations, significantly reducing the synthetic-to-real gap. Our pipeline enables researchers and companies to optimize and validate their volumetric capture setups virtually, ensuring robust performance upon real-world deployment.

Configurations

RegNet.

Different configurations for the VC setup: Mantis (exact digital twin), Cylinder (prioritize density) and Cube (prioritizing uniformity).

Quantitative Results

Quantitative.

Ablation studies

Ablations.

Qualitative results

Qualitative results.

Qualitative results for the five chosen characters using the chosen volumetric reconstruction methods on the Mantis configuration. All the methods perform very similarly, although they produce different kinds of very imperceptible artifacts. The last two columns are being compared to real-world scenes captured using a physical VC setup by Mantis Vision.

BibTeX

Coming soon