Matheus Souza

I am a Ph.D. student at KAUST Computational Imaging Group, working with Prof. Wolfgang Heidrich.

Email  /  Google Scholar  / 

profile photo
Education

  • 2022 - now  : Ph.D. in Computer Science, KAUST, Saudi Arabia.
  • 2020 - 2022: M.Sc. in Computer Science, KAUST, Saudi Arabia.
  • 2015 - 2020: B.Sc. in Electrical Engeneering, UFRN, Brazil.
  • Research

    My research is centered on the innovative fusion of deep reconstruction algorithms and end-to-end optics design. This involves developing deep learning models and formulating optimization strategies that incorporate optical elements into the computational loop. Focusing on the following topics:

  • Minimalistic cameras design.
  • Deep learning for optical design.
  • Computational cameras.
  • Optics-aware computational photography.
  • Latent Space Imaging
    Matheus Souza, Yidan Zheng, Kaizhang Kang, Yogeshwar Nath Mishra, Qiang Fu, Wolfgang Heidrich
    CVPR 2025. Arxiv pre-print (To be updated soon)

  • New paradigm for very low bandwidth image capture based on generative models latent space.
  • The demonstration of a range of downstream applications on this latent space with real hardware experimentation.
  • Limitations of Data-Driven Spectral Reconstruction - An Optics-Aware Analysis
    Qiang Fu*, Matheus Souza*, Suhyun Shin, Eunsue Choi, Seung-Hwan Baek, Wolfgang Heidrich
    Computational Optical Sensing and Imaging, 2024. Oral Presentation
    Full Paper Under Review. Arxiv pre-print

  • Comprehensive analysis of state-of-the-art data-driven hyperspectral imaging atypical overfitting.
  • Optical aberrations can provide encoding power if modeled correctly.
  • End-to-End Hybrid Refractive-Diffractive Lens Design with Differentiable Ray-Wave Model
    Xinge Yang, Matheus Souza, Kunyi Wang, Praneeth Chakravarthula, Qiang Fu, Wolfgang Heidrich
    Siggraph Asia 2024. Paper (Arxiv) / Paper (PDF) / Supp (PDF)

  • Differentiable ray-tracing and wave-propagation model.
  • End-to-End hybrid refractive-diffractive lenses design with prototypes.
  • MetaISP - Exploiting Global Scene Structure for Accurate Multi-Device Color Rendition
    Matheus Souza, Wolfgang Heidrich
    MetaISP VMV 2023 / Code
    CRISPnet: Color rendition ISP net. Paper (Arxiv)

  • We developed a model for learning multiple commercial ISPs.
  • Integrating global scene semantics, metadata information, and advanced deep learning techniques.
  • Collected synthetic and real-world datasets, consisting of RAW-RGB pairs from various devices.

  • The website template is from Dr. John Barron.