Denis Rozumny

Note: my name can also be spelled Denys Rozumnyi.

I am a Research Scientist at Meta Reality Labs. My main research interests include dynamic scene reconstruction, generative models, and photorealistic rendering.

I finished my PhD at the Computer Vision and Geometry Group, Department of Computer Science, ETH Zürich under supervision of Prof. Marc Pollefeys. I also worked closely with Prof. Martin Oswald, Prof. Vittorio Ferrari, and Prof. Jiri Matas.

Previously, I was a research intern at Google Research with Prof. Vittorio Ferrari and at Meta Reality Labs. Before that, I finished my MSc and BSc degrees from CTU in Prague, Center for Machine Perception under supervision of Prof. Jiri Matas.

My research interests are in 3D reconstruction of objects and scenes, video understanding, realistic rendering, object detection, tracking, 6D pose estimation. In particular, I focus on test-time optimization methods to solve those problems.

Email  /  CV  /  Google Scholar  /  Twitter  /  GitHub  /  LinkedIn

News

• Jun 2025: Two papers accepted to ICCV'25!
• Mar 2025: We released a paper on how to improve metrics for wireframe reconstructions.
• Nov 2024: I've joined Meta Reality Labs as a Research Scientist.
• Jul 2024: One paper accepted to ECCV'24: Master thesis student who I supervised.
• Jun 2024: Won 1st place in the Structured Semantic 3D Reconstruction challenge at CVPR'24 workshop USM3D.
• Mar 2024: Finished my research internship at Meta.

Research

We introduce an approach that takes advantage of generative models to correct errors and complete missing information in a Gaussian-based dynamic scene representation.

We introduce a recipe for generating immersive 3D worlds from a single image by framing the task as an in-context learning problem for 2D inpainting models.

Deblur-SLAM can successfully track the camera and reconstruct sharp maps for highly motion-blurred sequences. We directly model motion blur, which enables us to achieve high-quality reconstructions, both on challenging synthetic and real data.

We asked 3D modelers to rank wireframe reconstructions and compared it to the ranking by metrics. After I exploited issues in CVPR'24 challenge S23DR and won the 1st place, we came up with a better solution for the next challenge.

We leverage the available depth sensing signal on XR devices combined with self-supervision to learn a multi-modal pose estimation model capable of tracking full body motions in real time.

We propose a method for object retrieval in images that are affected by motion blur.

A diffusion-based model that deblurs and recovers shape of fast moving objects from a single image without a known background (for the first time).

We propose a novel method to produce generic 3D room layouts just from 2D segmentation masks, with which we annotate and publicly release 2246 3D room layouts on the RealEstate10k dataset.

We estimate 3D human poses from substantially blurred images, e.g. extension of Shape from Blur to SMPL human body model.

We propose to guide and improve 2D tracking with an explicit object representation, namely the textured 3D shape and 6DoF pose in each video frame.

A new algorithm for finding an unknown number of geometric models.

Extension of Shape from Blur to multiple frames with more complex trajectories and exposure time modeling.

The first method to estimate textured 3D shape and sub-frame 6D motion of fast moving objects from a single frame.

The first deep-learning based approach for fast moving object detection.

We propose DeFMO that given a single image with its estimated background outputs the object's appearance and position in a series of sub-frames as if captured by a high-speed camera (i.e. temporal super-resolution). This is the first deep-learning based approach for FMO deblurring.

Summarization and extension of our GCPR'19 (TbD-NC) and ICCVW'19 (TbD) papers.

We extend TbD pipeline to track fast moving objects in full 6 DoF, estimating simultaneously their 3D motion trajectory, 3D pose and object appearance changes with a time step that is a fraction of the video frame exposure time.

We apply post-processing with dynamic programming and curve fitting to obtain more accurate object trajectories.

We propose a novel approach called Tracking by Deblatting to track fast moving objects.

Our method learns sensor or algorithm properties jointly with semantic depth fusion and scene completion and can also be used as an expert system, eg to unify the strengths of various photometric stereo algorithms.

Introducing fast moving objects for the first time as objects that move over distances larger than their size in one video frame: new problem, new dataset, new metrics, new baseline.

We propose an automated method to detect, group and rectify arbitrarily-arranged coplanar repeated elements via energy minimization.

Supervising

• Francesco Girlanda: Gaussian Splatting SLAM with Deblurring Capabilities, Master thesis.
• Rong Zou: Retrieval Robust to Object Motion Blur, Master thesis, Nov 2023, accepted to ECCV'24.
• Yiming Zhao: Human from Blur: Human Pose Tracking from Blurry Images, Semester project, May 2023, accepted to ICCV'23.
• Rajat Thakur: Predicting 3D Shape and Texture of Fast Moving Cars, Semester project defended in March 2022.
• Adrian Klaeger: Temporal Super-Resolution of Multiple Fast-Moving Objects, Master thesis defended in September 2021, thesis, GitHub.
• Harish Rajagopal: Improving DeFMO With Learned Losses, Semester project defended in June 2021, report, GitHub.
• Julius Fricke: ADMM Algorithm Unrolling: Deblurring and Matting, Bachelor thesis defended in April 2021.

Reviewing

• Conferences: CVPR, ICCV, ECCV, NeurIPS, ICML, ICLR, 3DV, WACV.
• Journals: PAMI, IJCV.

Teaching

Adapted from here.