Proposal Title: XRTwinScape: Advancing Contextualized Learning and Training for Industry 5.0 Through Immersive Virtual Environments
Proposal Summary:
XRTwinScape is an innovative platform that uses digital twins—highly detailed virtual replicas of real-world workspaces—to provide immersive, risk-free training for industrial environments. This solution addresses the challenges of training in physical spaces, such as safety risks and logistical limitations, by offering realistic virtual environments for hands-on learning.
The project features two key components:
TwinScapeEditor – Allows easy creation of digital twins from simple photos and adds interactive XR training content.
TwinScapePlayer – Provides an immersive environment where trainees can interact with digital twins and complete task-based lessons.
Led by FMTS, CERICT, and PICA, the project combines expertise in industrial training, XR applications, and immersive environments. XRTwinScape will test its platform through three pilot use cases with 220 trainees:
Industrial Training – Simulating real work environments for safe, hands-on experience.
Problem Solving – Teaching workers to address complex, environment-specific issues.
Career Guidance – Exploring industrial roles through immersive experiences.
The platform’s advanced AI-powered rendering technology enables rapid, cost-effective digital twin creation, making training more accessible and scalable. It supports open-source collaboration, allowing adaptation across industries and ensuring long-term sustainability through a Software as a Service (SaaS) model.
XRTwinScape represents a transformative step in workforce training, aligning with Industry 5.0 by enhancing training effectiveness, accessibility, and future workforce preparedness.
Applicants:
Formamentis SPA Società Benefit
Centro Regionale Information Communication Technology – CeRICT scrl
Hi everyone, I am Alberto Accardo and I am a 3d modeler.
Today I will try to summarize the differences between two advanced techniques for generating models in a 3d digital space from images : photogrammetry and gaussian splatting.
Photogrammetry and gaussian splatting are two distinct approaches used in the creation and visualization of 3D models. Photogrammetry involves capturing multiple photographs of an object or scene from different angles, then employing complex algorithms to reconstruct a highly detailed and realistic 3D representation. This method excels in preserving intricate textures and accurate geometric details, making it particularly suitable for applications that demand high fidelity and precision. However, its reliance on numerous high-quality images and controlled lighting conditions can make the process both resource-intensive and time-consuming.
In contrast, gaussian splatting utilizes Gaussian functions to represent surfaces, effectively “splatting” overlapping blobs of data to form the overall 3D image. This technique offers a significant advantage in speed, rendering complex scenes quickly, which is highly beneficial for real-time applications and interactive environments. On the downside, gaussian splatting may not capture fine details as accurately as photogrammetry, potentially compromising the realism of the final model.
Thus, while photogrammetry is ideal for projects where detail and accuracy are paramount, gaussian splatting provides a more efficient alternative when rapid processing is required. The choice between these methods ultimately depends on the specific requirements of the project, including the balance between visual quality and computational efficiency. In summary, both techniques have their merits and limitations, making them complementary tools in the field of 3D modeling and visualization.
From a 3D modeler’s perspective, it’s important to consider not only the inherent technical differences but also how each technique integrates into your workflow. Photogrammetry can yield extremely realistic textures and precise geometry, but it often requires extensive post-processing—such as cleaning up meshes and optimizing textures—to make the models animation-ready or game engine compatible.
On the other hand, gaussian splatting excels in rapid visualization, making it useful for quick iterations and real-time previews during the design process. However, its lower level of detail might necessitate additional manual refinement or the use of hybrid techniques when the final output demands high fidelity.
I’m Andrea Amorosini, an AI engineer with a strong focus on 3D reconstruction and rendering. I’m currently developing a comprehensive Gaussian splatting pipeline—an innovative approach that utilizes Gaussian distributions to represent scenes with smooth blending and improved visual quality.
Unlike traditional methods that rely on point clouds or meshes, Gaussian splatting represents a scene using a collection of Gaussian distributions. Each Gaussian acts as a probabilistic primitive, capturing the uncertainty and natural variations within the scene. This approach results in smoother transitions and enhanced visual quality, making it a promising method for both static and dynamic 3D rendering.
Structure-from-Motion (SfM) Tools
The first stage of my pipeline involves recovering 3D structure from 2D images. For this, I’m exploring several advanced SfM solutions:
COLMAP:
Recognized as a gold standard in SfM, COLMAP excels in feature extraction, matching, and camera pose estimation. Its robust performance and extensive documentation make it a reliable choice for generating high-quality reconstructions.
GLOMap:
Emphasizing global optimization, GLOMap effectively handles diverse and challenging datasets. It ensures consistency and robustness in recovering the underlying 3D geometry of a scene.
Mast3r_sfm:
Designed to seamlessly integrate images from multiple sources, Mast3r_sfm is tuned to manage variations in image quality and perspective, making it suitable for large-scale scene reconstructions.
VGGsfm:
Developed by the Visual Geometry Group, VGGsfm offers a methodical framework for reconstructing 3D structures from unordered image collections, ensuring both accuracy and reliability.
Gaussian Splatting Tools
Once the SfM stage lays down a solid 3D foundation, I focus on the Gaussian splatting techniques, which enhance the rendering process:
Graphdeco Implementation:
This original approach employs graph-based optimization to efficiently manage complex scene structures, balancing computational efficiency with detailed representation.
Nerfstudio:
Integrating neural rendering with traditional geometric methods, Nerfstudio provides a versatile framework. It supports experimentation with neural radiance fields alongside Gaussian splatting, ideal for dynamic scene modeling.
Gsplat:
Gsplat is tailored for real-time performance, implementing advanced techniques to ensure efficient execution even in high-resolution settings. This tool is essential when both speed and visual quality are critical.
CF-3DGS:
By incorporating iterative refinement and feedback mechanisms, CF-3DGS enhances the accuracy and visual fidelity of the 3D models produced by the splatting process.
Opensplat:
Opensplat contributes additional capabilities to the Gaussian splatting process, further broadening the range of solutions available for high-quality scene representation, by focusing on performance and speed being the only implementation developed in C++.
I’m excited about how these advanced methodologies can reshape our approach to 3D reconstruction and rendering. My work is an ongoing journey of exploration and innovation, and I’m eager to exchange ideas with fellow researchers and practitioners.
Thank you for sharing your perspective as a 3D modeler! It’s really insightful to consider not just the technical differences between photogrammetry and Gaussian splatting, but also how they fit into a real workflow. Your breakdown it’s very insightful and thanks for this contribution! aaaaaand we can’t wait to see future update on your project, great to have you into the community.
Really impressive work you’re doing, and thanks for sharing - It’s always interesting to see innovative approaches like yours in 3D reconstruction and rendering.
It’ll be great to exchange ideas and dive deeper into these topics.
We’re eager to see the incredible things your projects will bring, and once again, a warm welcome to the community!
Hi Alberto, my name is Vlasis Kasapakis and I am an Assistant Professor working on XR applications for cultural heritage. Thanks for clearly explaining the differences between photogrammetry and gaussian splatting. I’ve experimented a bit with splatting for AR application development, but I’m also interested in exploring its potential for VR. For example, in the HERIT-ADAPT project, we used photogrammetry to reconstruct a cultural heritage site for a VR application. Based on your experience as a 3D modeler, in what types of projects or scenarios do you prefer photogrammetry, and when do you choose gaussian splatting? Are there cases where you combine both techniques?
Funny that your project name sounds like one of the software our company develops (https://www.ls-group.fr/xr-twin)
Thank you for the technical discussion on Gaussian Splatting and eager to discover more on your project !
Hello Vlasis!
Please accept my apologies for the delayed reply. Your message must have been lost among many emails.
First of all, congratulations on the HERIT-ADAPT project: the quality of the photogrammetry you have produced is truly remarkable.
However, I believe that the choice between photogrammetry and Gaussian Splatting should be driven primarily by the ultimate goal of your VR scenario. If the main requirement is direct interaction with objects (including the application of physical forces, collisions, and dynamic behaviors) photogrammetry is more suitable, as it generates fully editable 3D meshes on which you can define colliders and configure interaction logic. Conversely, while Gaussian Splatting offers a high level of visual realism, it does not provide true geometric surfaces on which to apply physics or the typical interactions of an immersive VR environment.
Of course, it is possible to use both techniques together in engines such as Unity or Unreal, subject to the usual constraints and compatibility considerations (lighting, shadows, physics, etc.). To leverage the strengths of both approaches, one could employ photogrammetry for the primary assets with which the user interacts (using meshes and colliders to handle physics) while applying Gaussian Splatting to vegetation, atmospheric details, or distant scenery, enriching the environment without overloading the interaction logic. This hybrid approach is particularly advantageous for creating high-fidelity digital twins, in museum or archaeological experiences where interactive objects coexist with extremely realistic reconstructed contexts, and in narrative walkthroughs that alternate between interactive areas and photographic panoramas, all while maintaining high standards of performance and immersion.
I remain available for any further clarification or exchange of thoughts on this topic!
