®®®® SIIA Público

Título del libro: International Conference On Artificial Intelligence, Computer, Data Sciences, And Applications, Acdsa 2026
Título del capítulo: Application of Neural Network Training for Lighting Estimation in Augmented Reality

Autores UNAM:
RODRIGO TERPAN ARENAS; JOSE LUIS PUNZO DIAZ; CELIA ANGELINA SANCHEZ PEREZ; JORGE ALBERTO MARQUEZ FLORES; ALFONSO GASTELUM STROZZI;
Autores externos:

Idioma:

Año de publicación:
2026
Palabras clave:

Comparative analyzes; Data augmentation; Deep learning; Digital Objects; Lighting estimation; Mixed reality; Neural networks trainings; Neural-networks; Real-world; Realistic Lighting


Resumen:

© 2026 IEEE.Lighting is an important component of Augmented Reality (AR) and Mixed Reality applications since it provides a level of realism to the digital objects. Part of the process of generating realistic lighting is the estimation of light in the real world. One of the approaches that is most compatible with consumer grade mobile devices is the use of neural networks using the input of the mobile device screens. We present a comparative analysis of neural network architectures to determine the best combination of variables, both synthetic and from the real-world, to train these architectures for the estimation of physical light direction in spherical Euler angles. Real life information was produced by using a custom-made testing environment that allowed for the consistent illumination of specific angles. Results indicate that real life input data is the most effective at producing accurate estimations, future work will improve upon this methodology and focus on ways of increasing the variability of the training database.


Entidades citadas de la UNAM: