I am a Ph.D. candidate at the School of Electrical Engineering, Tel-Aviv University, advised by Prof. Shai Avidan, and currently a Research Intern at the NVIDIA Research Israel AI Lab.
My research interests include AI, machine learning, and computer vision. More specifically, I work on 3D reconstruction methods such as NeRF and 3D Gaussian splatting, multi-view foundation models, and geometric correspondence in complex real-world scenes.
We introduce multi-view foundation models that equip powerful 2D vision backbones with the ability to reason jointly across multiple camera views. By explicitly incorporating camera geometry, our approach produces 3D-consistent semantic representations without per-scene optimization, enabling scalable 3D understanding and correspondence.
We address the problem of grounding partial, in-the-wild 3D reconstructions in a globally consistent reference model. By representing the reference scene as a feature-augmented 3D Gaussian Splatting model and solving an inverse feature-based optimization, our method recovers a global 6DoF pose and scale for each reconstruction — yielding consistent alignment even across disconnected scenes with no visual overlap.
We propose a frequency-aware decomposition of 3D Gaussian splatting that disentangles low- and high-frequency content in a scene. This separation improves rendering quality and robustness, and yields a more interpretable representation that is particularly useful for downstream tasks such as editing, cleanup, and reconstruction refinement.
We address the challenge of cleaning and improving neural 3D reconstructions in regions that are poorly observed or entirely unseen. Our method leverages semantic priors and multi-view consistency to hallucinate plausible geometry and appearance, yielding cleaner, more coherent reconstructions that generalize beyond the input views.
We introduce VF-NeRF, a viewshed-based approach for rigid NeRF registration. By modeling which parts of a scene are visible from each camera pose, our method robustly aligns NeRFs to reference geometry, even in the presence of partial observations and large-scale outdoor environments.
We propose a shape-consistent GAN framework for cross-modality adaptation between CT and MRI whole-heart scans. By enforcing geometric consistency across modalities, our method improves semantic segmentation performance when labeled data is limited or available only in one modality.