El jueves 15 de enero a las 16:00 en el Salón de actos del Edificio Ada Byron se realizará el acto de defensa de la tesis doctoral de Lorenzo Mur-Labadía, titulada "Learning visual models for egocentric perception".
El trabajo ha sido dirigido por los profesores de nuestro departamento, D. José J. Guerrero y D. Rubén Martínez-Cantín.
Resumen: Egocentric vision enables systems to perceive and reason about the world from the user’s perspective, requiring models that understand objects, affordances, and environments in dynamic first-person settings. This thesis advances egocentric perception through visual models of object affordances and spatial environments, together with their integration into multi-modal representations. We first model the world as a collection of functional objects by learning affordance-aware object part segmentation with uncertainty estimation. We then move beyond object-centric reasoning by introducing complementary environment representations, including multi-label affordance maps, Bayesian-robust semantic fusion, and implicit scene modeling based on decomposed neural feature fields. Building on these representations, we propose end-to-end architectures for short-term object interaction anticipation that ground predictions in past human behavior and environmental affordances. Finally, we extend egocentric perception through multi-modal alignment, addressing activity grounding in long untrimmed videos and cross-view object correspondence between first- and third-person perspectives. Overall, the proposed methods achieve state-of-the-art performance across multiple egocentric perception benchmarks and have been published in top-tier venues, including ICRA 2023, IROS 2023, ICCV 2023, ECCV 2024, CVPR 2025, ICCV 2025, CVIU 2025 and T-PAMI 2026.
Enlaces:
·
https://sites.google.com/unizar.es/lorenzo-mur-labadia/inicio
