Itai Lang

I am a Postdoctoral Researcher at the University of Chicago, working with Assistant Professor Rana Hanocka. I did my PhD in Electrical Engineering at Tel Aviv University with the guidance of Professor Shai Avidan. During my PhD, I interned at Google Research, where I worked with Michael Rubinstein.

I am a passionate researcher who seeks creative solutions to new problems and innovative applications. My research is focused on developing artificial intelligence for geometry processing, spanning the fields of computer vision, computer graphics, and machine learning.

Email  /  CV  /  Google Scholar  /  Github  /  Twitter

Geometry in Style produces an identity-preserving stylization of mesh geonetry by deforming the surface normals of the input shape.

MeshUp deforms a 3D mesh into multiple concepts specified by text, images, or other meshes, while enabling to control the strength and location of their manifestation.

We propose an interactive segmentation technique for 3D shapes that produces fine-grained customized segmentations based on user clicks.

We generate precise localizations and highly detailed local textures on 3D shapes using text guidance.

We build a procedural program with an interpretable parameter space to recover a high quality and easily manipulated mesh from a point cloud or a sketch input.

We synthesize colors over a 3D shape and use CLIP supervision to localize semantic regions using open vocabulary text prompts.

We use pure correspondence learning and direct refinement optimization to predict a highly accurate scene flow while using a fraction of the training data.

We apply a low frequency perturbation in the spectral shape domain to alter the reconstruction by a victim mesh autoencoder to a desired output geometry.

We use latent similarity and the point coordinates themselves to construct one point cloud by the other and achive an accurate dense matching with a small training data amount and without any correspondence supervision.

We perturb an input point cloud to attack an autoencoder model and change the reconstructed geometry to a different selected target shape.

We propose a differentiable relaxation to the sampling operation that enables learning a task-oriented sampling model in an end-to-end manner.

We propose a data-driven sampling approach for 3D point clouds that selects the most suitable subset of points for a downstream task.

This webpage was adapted from Jon Barron's webpage. We thank Jon Barron for sharing his source code.