Deep Generative Modelling in the Spatial Domain

Maxim Khomiakov: Learning to synthesize depth from sparse input

This project is about the modelling of 3D-structures with sparse input information. Our study will attempt to answer the question, how one may reconstruct buildings using partial digital input sources, such as imagery, LiDAR and vectorized polygons. The idea is to leverage measurements in areas of high quality to reconstruct 3D shapes in areas of limited or low quality image quality.

We shall focus on deep generative models. These mathematical models benefit from the ability to draw samples from a latent distribution, on which we assume data lies, as well as thriving in problems of high noise. Previous years have seen strong results in recent applications of these models in their ability to generate realistic imagery, image translation, in addition to the generation of 3D objects.

As many industries strive to digitize, the photovoltaic industry is no different. A number of challenges are still very much present. What is commonly referred to as solution design is an overly manual process. This revolves around the positioning of the panels, the sizing of the system, security needs for the electrician on the roof, in particular regarding wiring and materials. Without a good geometric representation of the building property, we are hindered in digitizing the solar acquisition process for the customer.

The aim is to alleviate these bottlenecks, by researching and developing methods that learn depth representations of urban architecture using partial input data sources.

PhD project

By: Maxim Khomiakov

Section: Cognitive Systems

Principal supervisor: Jes Frellsen

Co-supervisors: Michael Riis Andersen, Simen Fure Jørgensen

Project titleDeep Generative Modelling in the Spatial Domain

Term: 01/08/2020 → 24/01/2023


Maxim Khomiakov
External PhD student
DTU Compute


Jes Frellsen
Associate Professor
DTU Compute
+45 45 25 39 23


Michael Riis Andersen
Associate Professor
DTU Compute
+45 45 25 34 09