Safa C. Medin

I am a final year PhD student at MIT EECS, supervised by Prof. Gregory W. Wornell and a Student Researcher at Google AR & VR, supervised by Abhi Meka and Thabo Beeler. I am interested in synthesis, animation, and manipulation of 3D digital humans for applications in extended reality 🥽. Previously, I worked on various problems in computational imaging, particularly classifying and reconstructing scenes in non-line-of-sight settings. Please feel free to reach out if you would like to chat!

  News

  • May 2024 Our work FaceFolds wins the Best Student Paper Award at I3D 2024! 🏆

  • Mar 2024 Our paper, "FaceFolds: Meshed Radiance Manifolds for Efficient Volumetric Rendering of Dynamic Faces", is accepted to I3D 2024.  | Paper | Webpage | Video

  • Jan 2024  I started my third internship at Google AR & VR in Cambridge, MA.

  • Dec 2023  I am now officially a PhD candidate!

  • Jan 2023  I started my second internship at Google AR & VR in Cambridge, MA.

  • Aug 2022  Our paper, "Can Shadows Reveal Biometric Information?", is accepted to WACV 2023.  | Paper

  • May 2022  I started my summer internship at Google AR & VR in San Francisco, CA. 🌁 🌉  | Photo

  • Dec 2021  Our paper, "MOST-GAN: 3D Morphable StyleGAN for Disentangled Face Image Manipulation", is accepted to AAAI 2022.  | Paper | Video

  • Sep 2021  I received my Master's degree from MIT EECS! 🎓 | Thesis

  • Jul 2021  Our paper, "Identity-Expression Ambiguity in 3D Morphable Face Models", is accepted to FG 2021.  | Paper

  • Jun 2021  I started my second internship at MERL.

  • May 2020  I started my summer internship at MERL.

  • Sep 2019  I started my PhD at MIT EECS! 🏛️ 🦫


Research

Efficient Volumetric Rendering of Photoreal Avatars

Volumetric rendering of dynamic face captures is a challenging problem with applications in extended reality, such as immersive communication and 3D telepresence. While recent methods have achieved high-quality and photorealistic 3D rendering of human head avatars, they often come with low rendering efficiency or high storage requirements. This project is aimed at developing novel 3D representations that run natively on legacy renderers without any custom neural network integration, while also allowing for a graceful trade-off between image quality and rendering efficiency via standard graphics operations at inference time.

Unbiased and Diverse 3D Face Reconstruction

Single-view 3D face reconstruction is an ill-posed problem and many of its applications are expected to work in complex environments with uncertainties that are challenging to model. This project is aimed at leveraging pretrained diffusion models to guide the 3D reconstruction process and output a diverse set of solutions to identify biases in the datasets as well as reconstruction algorithms, and subsequently mitigate these biases.

3D-Aware Face Image Manipulation

Advances in generative adversarial networks (GANs) have led to remarkable achievements in face image synthesis. While typical GAN-based methods have the ability to generate photorealistic face images, it is often difficult to manipulate the characteristics of the generated faces in a 3D-controllable and disentangled manner. In this project, our goal is to integrate a 3D face model into the photorealistic image generation process and achieve a fully disentangled face image manipulation pipeline that allows for extrapolating beyond the variations in the datasets.

Occluder-aided Non-Line-of-Sight Imaging

Non-line-of-sight (NLOS) imaging is the study of extracting information from hidden scenes based on visible scenes that are in our direct line-of-sight. A popular NLOS imaging setting, occluder-aided imaging, exploits occluding structures in the scenes to achieve this task. In this project, we aim to develop learning-based methods for a variety of occluder-aided imaging applications to achieve reliable NLOS imaging systems as well as identify whether or not seemingly benign optical phenomena can leak sensitive information about the hidden scenes.

Bernoulli Parameter Estimation in Active Imaging

In active imaging systems, periodic illumination pulses sent from the source can either be absorbed by the scene or reflect from it. Representing the probability of reflection from each patch as a Bernoulli parameter, the image acquisition process can be modeled as a problem of estimating arrays of Bernoulli parameters. In this setting, varying resources across multiple patches can yield significant improvements in acquisition efficiency. Motivated by this, we aim to develop adaptive acquisition strategies that achieve such performance improvements.


Publications

  • Safa C. Medin, Gengyan Li, Ruofei Du, Stephan Garbin, Philip Davidson, Gregory W. Wornell, Thabo Beeler, and Abhimitra Meka, "FaceFolds: Meshed Radiance Manifolds for Efficient Volumetric Rendering of Dynamic Faces", in Proceedings of the ACM in Computer Graphics and Interactive Techniques, 2024.   | Paper | Webpage | Video

  • Safa C. Medin, Amir Weiss, Frédo Durand, William T. Freeman, and Gregory W. Wornell, "Can Shadows Reveal Biometric Information?", in IEEE/CVF Winter Conference on Applications of Computer Vision, 2023.   | Paper

  • Safa C. Medin, Bernhard Egger, Anoop Cherian, Ye Wang, Joshua B. Tenenbaum, Xiaoming Liu, and Tim K. Marks, "MOST-GAN: 3D Morphable StyleGAN for Disentangled Face Image Manipulation", in 36th AAAI Conference on Artificial Intelligence, 2022.   | Paper | Video

  • Bernhard Egger, Skylar Sutherland, Safa C. Medin, and Joshua B. Tenenbaum, "Identity-Expression Ambiguity in 3D Morphable Face Models", in 16th IEEE International Conference on Automatic Face and Gesture Recognition, 2021.   | Paper

  • Safa C. Medin, "Learning-based Methods for Occluder-aided Non-Line-of-Sight Imaging", Master's Thesis, Massachusetts Institute of Technology, 2021.   | Thesis

  • Safa C. Medin, John Murray-Bruce, David Castañón, and Vivek K. Goyal, "Beyond Binomial and Negative Binomial: Adaptation in Bernoulli Parameter Estimation", in IEEE Transactions on Computational Imaging, vol. 5, no. 4, pp. 570-584, Dec. 2019.   | Paper

  • Safa C. Medin, John Murray-Bruce, and Vivek K. Goyal, "Optimal Stopping Times for Estimating Bernoulli Parameters with Applications to Active Imaging", in IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2018.   | Paper