Ivaylo 'Ivo' Boyadzhiev

I am a Senior Applied Science Manager at Zillow Group, where currently I lead a team of Applied Scientists and Machine Learning Engineers working on indoor scene reconstruction and understanding, using scalable off-the-shelf hardware solutions, like smart phones and 360 RGB cameras. I also co-managed the release of our open-source, free for academic research, Zillow Indoor Dataset, published at CVPR 2021.

At Zillow I've worked on Floor Plan Reconstruction for Interactive Floor Plans, Panorama Stitching for 3D Home®, Improving Video Stabilization for Video Walkthrough, Intrinsic Image Decomposition for Zillow Digs. I did my PhD at Cornell University, where I was advised by Kavita Bala and I spent 3 summers 2012-2014 working at Adobe Research under the supervision of Sylvain Paris. My PhD thesis centered around the problems of designing algorithms to assist advanced computational photography tasks, such as lighting design, white balance under mixed lighting, and image filters for material editing.

Email  /  CV  /  Google Scholar

I'm interested in computer vision, machine learning, computational photorgraphy and image processing. Much of my recent research is about indoor reconstruction, scene understanding and wide-baseline camera localization in challenging interior settings.

We present CoVisPose, a new end-to-end supervised learning method for relative camera pose estimation in wide baseline 360° indoor panoramas.

We propose a new system for automatic 2D floorplan reconstruction that is enabled by SALVe, our novel pairwise learned alignment verifier. The inputs to our system are sparsely located 360° panoramas, whose semantic features (windows, doors, and openings) are inferred and used to hypothesize pairwise room adjacency or overlap.

We describe a novel, weakly-supervised, approach to decompose a single panorama of an empty indoor environment into four appearance components: specular, direct sunlight, diffuse and diffuse ambient without direct sunlight.

We propose to extract topological information from room attributes using an enhancement of GNNs that we call Iterative and adaptive graph Topology Learning (ITL). ITL progressively predicts multiple relations between rooms; at each iteration, it improves node embeddings, which in turn facilitates the generation of a better topological graph structure.

We address the problem of multimodal pose estimation, given an RGB image and a 2D schematifc floor-plan. We introduce the concept of latent space rendering, where 2D pose hypotheses on the floor map are directly rendered into a geometrically-structured latent space by aggregating viewing ray features.

Given two-view panoramas, we propose the first network to jointly estimate the complete room layout and refine a given coarse relative pose between them.

The Zillow Indoor Dataset (ZInD) is available for download (academic research only)! It features more than 70k omnidirectional panoramas from more than 1.5k unfurnished homes, and has room layout and annotation (windows and doors) information.

We address challenging cases for SfM (in indoor settings) by fusing of single and multi-view reconstruction techniques via geometric relationship detection (e.g., detecting 2D lines as coplanar in 3D).

We propose a simple do-it-yourself (DIY) setup with a video camera on a tripod recording the product while the user waves an area light source around it for a couple of minutes.

We propose an extended family of band-sifting filters and study their perceptual effects for image-based material editing.

We propose a novel user-guided lighting design method based on capturing several photos from a fixed viewpoint with a moving light source.

We propose a solution to the ill-posed mixed light white balance problem, based on relative reflectance properties, that are easy for humans to provide in the form of scribbles.

Design and source code from Jon Barron's website