Speakers
- Bihan Wen — Nanyang Technological University
Abstract
Modern computational imaging systems capture rich visual information across diverse modalities and scales, yet translating raw measurements into accurate, interpretable representations remains a fundamental challenge. This talk presents our recent advances in physics-driven machine learning for computational imaging, spanning three interconnected themes: reconstruction, analysis, and generation. In the reconstruction thread, we explore novel view synthesis, 3D scene understanding, and high-fidelity sparse-view reconstruction. In the analysis thread, we discuss how to leverage physical prior and how foundation models can be grounded with physical constraints for robust generalisation. In the generation thread, we present physics-aware approaches for reflection removal and shadow removal that disentangle physically distinct image formation components. Across all themes, a unifying philosophy emerges: embedding domain knowledge of underlying optical physics into the learning framework yields models that are more accurate, data-efficient, and interpretable.
Looking for more in the field?
Explore more events from JIVP Webinar Series and Springer Nature's Computational Science.