Ali Osman Ulusoy

I am a software engineer at Google in Seattle where I work on Project Starline. My background is in computer vision.

Before joining Google, I worked at Microsoft on HoloLens and Azure Spatial Anchors. I did my Phd in electrical engineering with Joseph Mundy at Brown University. After Phd, I did a post-doc at the Max-Planck Institute for Intelligent Systems in Tubingen where I was advised by Michael J Black and Andreas Geiger.

Email  /  Google Scholar  /  LinkedIn  /  CV

My research interests span computer vision, graphics and machine learning. I'm particulary interested in inferring 3D structure and motion from images. Representative papers are highlighted below.

Neural networks can learn to parse a 3D model into a small set of Superquadrics.

RayNet is a neural network for 3D reconstruction from images. We integrate unrolled belief propagation within the network, so that it doesn't have to learn about geometry of perspective projection and occlusion, but rather focus on learning view-invariant features.

Learning depth-fusion can be much more accurate than traditional TSDF fusion and TV-L1 fusion.

We accelerate 3D convolutional neural networks using octrees.

We extend our 3DV 2015 paper with object-level 3D shape priors. Inference in this factor graph yields dense 3D geometry in addition to the 6DOF pose of objects in it.

We extend Horn and Schunck's optical flow formulation to 3D volumetric models and use it for compression.

We extend our 3DV 2015 paper with non-local planarity surface priors. Much of the paper is about efficient inference in the factor graph.

We formulate multi-view stereo as inference in a factor graph. This allows exposing the inherent ambiguities in surface reconstruction from images, and incorporating surface priors.

We use a stream of images captured over time to continuously update a 3D model with changes in the environment.

We study the problem of registering volumetric 3D models, by evaluating sensitivity to discretization, camera registration errors, and changes in illumination.

We reconstruct arbitrary dynamic 3D scenes accurately using a view-dependent volumetric modeling, and efficiently by exploiting redundancies in space and time.

We solve 3D object recognition in the context of volumetric 3D models.

We fuse images from different viewpoints into a volumetric probabilistic model and extract a textured mesh from it.

We use factors graphs and belief propagation to solve 3D scanning using structured light.

We project a grid pattern with unique spacings onto an object to reconstruct its 3D structure.