Yehonathan Litman

I'm a fourth year PhD student at in the Robotics Institute, Carnegie Mellon University, advised by Shubham Tulsiani and Fernando De la Torre. I have previously interned at Meta Reality Labs and was a visiting researcher at FAIR.

I'm focusing on how large generative models can serve as priors for faithfully reconstructing the world by capturing its underlying intrinsic properties. My work is supported in part by an NSF Graduate Research Fellowship.

I am always open to talking about research ideas or potential collaborations. Feel free to email me.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

Research

An efficient video inpainting control framework that focuses computation only where it is needed, achieving 10x compute efficiency over state-of-the-art generative editing methods while improving editing quality.

A compact, coarse-to-fine neural material model using texture-space shading and spatiotemporal computation amortization, achieving over 90 FPS on Meta Quest 3 with visual quality comparable to NeuMIP.

Efficiently relight any number of input images to a target lighting using inferred intrinsic property cues, extensible to 3D novel view relighting.

Distill a material diffusion prior model into material parameters during inverse rendering to improve intrinsic decomposition and novel relighting.

Reconstruct input images in a 3D occupancy grid and project to bird's eye view of the environment, register that to satelite imagery data to correct SLAM drift in real time without requiring GPS.

Dynamically assign a drone agent leader to heuristically assemble a large variety of drone swarm formations using only vision and onboard sensors.

Retain features across frames selectively based on shared information and use hardware acceleration to dramatically improve VIO SLAM efficiency.

Addressing the visual observability problem in VIO by using invariant filtering to improve efficiency and accuracy.

Awards

Misc.

Credits: Jon Barron.