Dor Verbin

I'm a research scientist at Google DeepMind in San Francisco. Before that I got my Ph.D. in computer science from the Harvard School of Engineering and Applied Sciences, where I was advised by Todd Zickler.

Previously, I received a double B.Sc. in physics and in electrical engineering from Tel Aviv University, after which I spent a couple of years as a researcher at Camerai (acquired by Apple), working on real-time computer vision algorithms for mobile devices.

Email  /  Google Scholar  /  Twitter  /  Github

I am interested in computer vision and graphics. More specifically, I enjoy working on methods that use domain-specific knowledge of geometry, graphics, and physics, to improve computer vision systems.

Casting reflection cones inside NeRF lets us synthesize photorealistic specularities in real-world scenes.

Using radiance caches, importance sampling, and control variates helps reduce bias in inverse rendering, resulting in better estimates of geometry, materials, and lighting.

We do 3D relighting by sampling from a single-image relighting diffusion model, and distilling it into a latent-variable NeRF.

Shadows cast by unobserved occluders provide a cue for recovering materials and high-frequency illumination from images of diffuse objects.

We finetune an image diffusion model to accept multiview inputs, then use it to regularize radiance field reconstruction.

We use a generative text-to-image model for creating videos with extreme zoom ins.

Applying anti-aliasing to a discrete opacity grid lets you render a hard representation into a soft image, and this enables highly-detailed mesh recovery.

We design a new boundary-aware attention mechanism to quickly find boundaries at images with low SNRs. The output is similar to the field of junctions, but this works ~100x faster!

We use neural fields to recover editable UV mappings for challenging geometry (e.g. volumetric representations like NeRF or DreamFusion, or meshes extracted from them).

We combine mip-NeRF 360 and Instant NGP to reconstruct very large scenes.

We treat each point in space as an infinitesimal volumetric surface element. Using MC-based rendering to fit this representation to a collection of images yields accurate geometry, materials, and illumination.

We achieve real-time view synthesis by baking a high quality mesh and fine-tuning a lightweight appearance model on top.

We achieve real-time view synthesis using a volumetric rendering model with a compact representation combining a low resolution 3D feature grid and high resolution 2D feature planes.

We modify NeRF's representation of view-dependent appearance to improve its representation of specular appearance, and recover accurate surface normals. Our method also enables view-consistent scene editing.

We extend mip-NeRF to produce realistic results on unbounded scenes.

By modeling each patch in an image as a generalized junction, our model uses concurrencies between different boundary elements such as junctions, corners, and edges, and manages to extract boundary structure from extremely noisy images where previous methods fail.

We present a simple condition for the uniqueness of a solution to the shape from texture problem. We show that in the general case four views of a cyclostationary texture satisfy this condition and are therefore sufficient to uniquely determine shape.

We formulate the shape from texture problem as a 3-player game. This game simultaneously estimates the underlying flat texture and object shape, and it succeeds for a large variety of texture types.

I'm also using Jon's website template.