By Benjamin Chiem

In an era where critical infrastructure, from power grids to transport networks, demands unprecedented levels of monitoring, we wanted to build a solution that would change the status quo for companies across the sector. As a collaborative venture between a team of AI experts from Euranova, and Stemme Belgium, a leader in high-performance aerial platforms, the Eagle Eye project is setting a new standard for 3D Digital Twins.

Our mission is clear: to transform raw aerial imagery into high-fidelity, photorealistic 3D models that are not just visually compelling, but geometrically precise and semantically intelligent. Over the past year, our research has moved from theoretical frameworks to a robust, cloud-native production pipeline, addressing the complex issues at the intersection of geospatial engineering and computer vision. Let’s dive into the journey of our engineers.

[From aerial images (left), our advanced reconstruction pipeline generates a georeferenced (center) and precise 3D model (right)]

The Stakes: Why 3D Intelligence Matters

Traditional infrastructure inspection is often slow, hazardous, and limited by the quality of 2D documentation. While 3D reconstruction is not new, "standard" solutions frequently fail when faced with the specificities of industrial needs:

• Complex Geometries: Thin structures like power line pylons and cables are notoriously difficult for standard algorithms to reconstruct, often appearing as "melted" or blurred artifacts.
• Aerial Constraints: To cover large areas efficiently, aerial platforms often capture images from high altitudes, resulting primarily in "nadir" (top-down) views. This operational constraint leads to a lack of varied angles (parallax), which typically degrades 3D reconstruction quality.
• Massive Data Volumes: Capturing the detail required for true inspection-grade digital twins generates massive, ultra-high-resolution aerial datasets. Processing this data volume efficiently requires more than just high-performance computing; it demands a carefully engineered, scalable architecture.

[Even with high-altitude imagery, Eagle Eye enables the remote inspection of critical infrastructure such as electrical pylons.]

Technical Innovations: Science in Service of Industry

To address these complex requirements, a scalable, strong, and unified software codebase was essential. Here are the four technical levers we deployed:

1. 3D Gaussian Splatting

Moving beyond traditional photogrammetry, we have adopted 3D Gaussian Splatting. Unlike standard point clouds, this method represents the world as millions of optimized 3D particles (Gaussians). This technique appeared in 2023 and benefits from continuous rapid developments in the scientific community, which we actively integrate to maintain a state-of-the-art solution. Our rigorous benchmarks led us to a reference pipeline that delivers superior visual quality, significantly faster reconstruction times, and ensures a manageable model size even for the reconstruction of large outdoor areas.

2. Focused Densification: Precision Where it Counts

One of our key features in development is Focused Densification. By integrating SAM2 (Segment Anything Model 2), we allow users to define Regions of Interest (ROIs) directly from 2D images. Our algorithm then concentrates its computational power on these areas.

• The result: Our approach yielded a 16-fold increase in Gaussian density for critical structures. This allowed for the reconstruction of fine details, enhancing visual fidelity in vital areas without losing the broader context of the reconstruction. These advancements are achieved while maintaining streamlined model dimensions and optimized computational usage.

[Focused densification is a key feature developed by the team of experts at Euranova. Based on the user preference, our algorithm increases the details on these areas of interest]

3. AI-Driven Depth Supervision

By leveraging AI foundation models, we generate relative depth maps for every frame to address the "nadir view" challenge we mentioned earlier. This approach provides a geometric guidance for our 3D models, transforming our geometric consistency correlation from a marginal 0.109 to an impressive 0.772. This ensures all structures are accurately anchored in 3D space rather than "floating" in the reconstruction.

4. Semantic Intelligence

We aren't just building models to be looked at; we are building models to be queried. We embed semantic data directly into the 3D scene, transforming the digital twin into an interactive database. This allows users to conduct natural language queries, receiving instant, precise responses in 3D space.

[Interact with your 3D model: using open-vocabulary queries, detect and highlight key features in your digital twin]

Performance and Scalability

To efficiently process the massive, high-resolution aerial datasets required for true inspection-grade digital twins, we have engineered a custom, cloud-native architecture. This platform moves beyond the rigidities of traditional on-premise solutions and leverages dynamic scalability to handle fluctuating computational loads.

Agile, Cloud-Native Architecture for Continuous Innovation

Our custom, cloud-native architecture is built for maximum agility and power. It operates on a strict separation of concerns, decoupling the smart control layer (orchestration) from on-demand computational workers. By isolating every algorithmic step using Docker containers, we ensure complete experiment reproducibility and prevent conflicts. This modularity is key: it allows us to systematically benchmark and integrate new 3D reconstruction methods, from the latest research breakthroughs to our own proprietary solutions, with speed and confidence. The entire process is managed by Dagster, a modern orchestrator that provides crucial data traceability, guaranteeing the integrity of every result.

Conclusion and Future Perspectives

In its first year, Eagle Eye has grown from a research project into a powerful platform that’s raising the bar for aerial intelligence. Moving forward, we want to turn these high-quality digital twins into tools that drive real-world decisions. We’re focused on integrating our tech smoothly with modern enterprise GIS systems, building automated tools to spot infrastructure risks before they become problems, and staying at the leading edge of 3D vision research. We’re turning Eagle Eye into a whole new way to manage critical assets, bringing tomorrow’s infrastructure intelligence to the world today.

Financial Support

The project is supported by European Union and Région wallonne through the European Regional Development Fund. For more information regarding the Structural Funds and supported initiatives, please visit the official Walloon Government websites:

• europe.wallonie.be
• enmieux.be