Pratul Srinivasan

I'm a research scientist at Google DeepMind in San Francisco, where I work on problems at the intersection of computer vision, computer graphics, and machine learning.

I received my PhD from the EECS Department at UC Berkeley in December 2020, where I was advised by Ren Ng and Ravi Ramamoorthi.

During my PhD, I interned twice at Google Research: at Mountain View in 2017 (hosted by Jon Barron in Marc Levoy's group) and at New York City in 2018 (hosted by Noah Snavely).

I graduated from Duke University in 2014, where I worked with Sina Farsiu on research problems in medical computer vision.

Email  /  CV  /  Google Scholar  /  Twitter

Research and Publications

* denotes equal contribution co-authorship

3D relighting by distilling samples from a 2D image relighting diffusion model into a latent-variable NeRF.

Use neural fields to recover editable UV mappings for challenging geometry (e.g. NeRFs, marching cubes meshes, DreamFusion).

A system built around diffusion and NeRF that does text-to-3D, image-to-3D, and few-view reconstruction, trains in 1 minute, and renders at 60FPS in a browser.

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.

Shadows cast by unobserved occluders provide a high-frequency cue for recovering illumination and materials.

Using a multi-image diffusion model as a regularizer lets you recover high-quality radiance fields from just a handful of images.

Use a text-to-image model to generate consistent content across drastically varying scales.

Combining mip-NeRF 360 and Instant NGP lets us reconstruct huge scenes.

We use SDFs to bake a NeRF-like model into a high quality mesh and do real-time view synthesis.

We use volumetric rendering with a sparse 3D feature grid and 2D feature planes to do real-time view synthesis.

ViT-VQGAN plus a NeRF-based decoder that enables both single-image view synthesis and 3D generation.

Construct a personalized 3D model from an unstructed photo collection.

We apply ideas from NeRF to the problem of reconstructing the dynamic emissive volume around a black hole.

We fix NeRF's shortcomings when representing shiny materials, greatly improve NeRF's normal vectors, and enable intuitive material editing.

We build city-scale scenes from many NeRFs, trained using millions of images.

Free-viewpoint rendering of any body pose from a monocular video of a human.

We train NeRFs directly on linear raw camera images, enabling new HDR view synthesis applications and greatly increasing robustness to camera noise.

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

We incorporate lidar data and explicitly model the sky to reconstruct urban environments with NeRF.

We apply dense depth completion techniques to freely-available sparse stereo data to guide NeRF reconstructions from few input images.

We recover relightable NeRF-like models from images under a single unknown lighting condition.

We modify NeRF to output volume density and emitted radiance at a volume of space instead of a single point to fix NeRF's issues with sampling and aliasing.

We "bake" a trained NeRF into a sparse voxel grid of colors and features in order to render it in real-time.

We deblur dual-pixel images by representing the scene as a multiplane image and carefully considering dual-pixel optics in an optimization framework.

We recover relightable NeRF-like models using neural approximations of expensive visibility integrals, so we can simulate complex volumetric light transport during training.

We use meta-learning to find weight initializations for coordinate-based MLPs that allow them to converge faster and generalize better.

Training a network that blends source views using a NeRF-like continuous neural volumetric representation, for NeRF-like performance without per-scene training.

Mapping input coordinates with simple Fourier features before passing them to a fully-connected network enables the network to learn much higher-frequency functions.

We recover relightable NeRF-like models by predicting per-location BRDFs and surface normals, and marching light rays through the NeRF volume to compute visibility.

We optimize a simple neural network to represent a scene as a 5D function (3D volume + 2D view direction) from just a set of images, and synthesize photorealistic novel views.

We represent scenes as multi-layer panoramas with depth for VR view synthesis.

We predict a multiscale light volume from an input stereo pair, and render this volume to compute illumination at any 3D point for relighting inserted virtual objects.

We develop a deep learning method for rendering novel views of complex real world scenes from a small number of images, and analyze it with light field sampling theory.

We use Fourier theory to show the limits of view extrapolation with multiplane images, and develop a deep learning pipeline with 3D inpainting for better view extrapolation results.

We train a neural network to estimate a depth map from a single image using only images with different-sized apertures as supervision, and use this to synthesize artificial bokeh.

We show that properly considering the eye's aberrations when rendering for VR displays increases perceived realism and helps drive accomodation.

We train a neural network to predict ray depths and RGB colors for a local light field around a single input image.

We develop Fourier theory to describe the effects of camera motion on light fields, and an optimization algorithm for deblurring light fields captured with unknown camera motion.

We develop a 4D light field descriptor and an algorithm to use these to compute scene flow (3D motion of observed points) from two captured light fields.

We develop an algorithm that jointly considers cues from defocus, correspondence, and shading to estimate better depths from a light field.

We develop a classification algorithm to detect diseases from OCT images of the retina.

We develop a segmentation algorithm to quantify the shape of retinal layers in OCT images that is robust to deformations due to disease.

You've probably seen this website template before, thanks to Jon Barron.
Last updated December 2023.