Tzofi Klinghoffer

Research Assistant
  • Camera Culture

I am a PhD student studying computer vision and machine learning at MIT Media Lab, where I am advised by Prof Ramesh Raskar. My interests lie broadly in the intersection of computer vision, computational imaging, and graphics. Topics of interest include 3D vision, neural simulation (e.g. neural fields), and data-driven imaging.

Previously, I was a software engineer in the Alexa AI group at Amazon, and, before that, an AI researcher at MIT Lincoln Laboratory, where I studied computer vision and unsupervised representation learning, and, in particular, their applications to medical images and human health. Since starting my PhD, I have spent summers at NVIDIA and Meta Reality Labs.

For more up-to-date information about recent work, news, and publications, please visit my personal website. If you are interested in collaboration, please email me.

Selected Publications:

- PlatoNeRF: 3D Reconstruction in Plato's Cave via Single-View Two-Bounce Lidar (CVPR 2024) - we model two-bounce light transport with NeRF, using lidar for supervision. Collaboration with Meta AI.

- DISeR: Designing Imaging Systems with Reinforcement Learning (ICCV 2023) - we define a search space for imaging systems and use reinforcement learning to search it, while jointly optimizing a perception model for a downstream task.

Towards Viewpoint Robustness in Bird's Eye View Segmentation (ICCV 2023) - we use novel view synthesis to enable autonomous vehicle perception models to generalize to diverse camera rigs. Collaboration with NVIDIA Research.

ORCa: Glossy Objects as Radiance-Field Cameras (CVPR 2023) - we use reflections to recover images and depth maps of parts of scenes that are outside the field of view of the camera.

Physics vs. Learned Priors: Rethinking Camera and Algorithm Design for Task-Specific Imaging (ICCP 2022) - we provide a framework and perspective on past work, current challenges, and future directions for task-specific camera design using machine learning.

Towards Learning Neural Representations from Shadows (ECCV 2022) - can shadows alone be used to recover the 3D shape of objects? We show that this is possible using neural radiance fields.

Physically Disentangled Representations (ECCV Workshops 2022) - we provide a method for using de-rendering for general purpose visual representation learning, showing it outperforms many other methods for generative representation learning for classification tasks.


I have served as a reviewer for conferences and workshops, including CVPR (2x), ECCV, CVPR Workshop on Autonomous Driving (WAD), and ECCV VOLI.