Generative AI is now capable of producing highly realistic 3D assets with unprecedented quality. But beyond content creation, can AI also improve established methods for 3D model digitization, such as photogrammetry?
At DSP 2025 (https://2025.ic-dsp.org), researchers from CNR presented a paper demonstrating how AI can do just that. By introducing a carefully designed synthetic training dataset and a neural network with a custom loss function, they were able to significantly reduce specular highlights in photographs. This enhancement leads to more accurate and visually consistent photogrammetric 3D reconstructions.
What is particularly notable is that this AI-driven improvement integrates seamlessly into the traditional photogrammetry pipeline. By operating between the image acquisition and reconstruction stages, the method preserves the standard workflow while boosting the final model quality.
Reference:Â
Marco Callieri, Massimiliano Corsini, Somnath Dutta, Daniela Giorgi, Marco Sorrenti:
AI-driven specular removal for 3D asset creation. Proceedings DSP 2025, 25th International Conference on Digital Signal Processing, Special Session on Digital Twins and XR: Signal Processing Challenges and Emerging Technologies