Oleg Tolochko

Technical University of Munich¹University of Cambridge²

Oleg wearing a blue IKEA bucket hat that covers one eye — **Figure 1.** Left: the author. Right: also the author.

Abstract

I'm a master's student in Informatics at TU Munich, working on 3D computer vision.

Right now I'm doing research on point cloud shape completion with the Computer Vision for Digital Twins group at the University of Cambridge. Before that I worked on uncertainty quantification at the Chair of Artificial Intelligence and Machine Learning at LMU Munich.

1. Publications

Nothing published yet.

2. Research and projects

APEX: Anchored Convex Polytopes for Explicit Shape Completion

University of Cambridge, CV4DT × TUM Data Innovation Lab, 2026 – present · Paper in preparation

Shape completion from a partial point cloud directly to a mesh, with the completed shape represented as a union of anchored convex polytopes. Supervised by Dr. Olaf Wysocki.

Three novel views rendered from a reconstructed university corridor.

Long-scene Gaussian splatting with non-rigid chunk alignment

TU Munich, Machine Learning for 3D Geometry, 2026

Reconstructs long image sequences in chunks. Each chunk is a 3D Gaussian Splatting scene initialised from DA3-Streaming, and neighbouring chunks are aligned with Sim(3) transforms and a non-rigid deformation field. Evaluated on ScanNet++.

Completed ShapeNet point clouds: airplane, bench, car, display, knife and lamp.

Zero-shot point cloud completion with generative priors

TU Munich, Seminar on Agents for 3D Generative AI, 2026

Technical report on GenPC (CVPR 2025), which completes a partial scan by rendering it to a depth prompt, generating an image and lifting that image to 3D. I re-ran the pipeline with newer foundation models (Qwen-Image-Edit, GPT Image 2, TRELLIS.2, Pixal3D) on real and ShapeNet scans.

ERA5 temperature fields with an injected anomaly, next to the epistemic and aleatoric uncertainty maps.

Bayesian hypernetworks for epistemic and aleatoric uncertainty

LMU Munich, Chair of AI and Machine Learning, 2026

Extends HyperDM (NeurIPS 2024), which separates epistemic and aleatoric uncertainty with a single diffusion model. Adds a Bayesian hypernetwork variant, MC-dropout baselines, a sparse-view CT experiment on LUNA16 and anomaly localisation benchmarks on ERA5.

[code]

Illustration: ten ensemble predictions scattered on a three-class probability simplex around a soft label.

Learning from ambiguous labels in Probly

LMU Munich, Chair of AI and Machine Learning, 2026

Experiment setup for the open-source uncertainty library Probly: deep ensembles trained on soft-label data, aleatoric/epistemic decomposition, Entmax and conformal evaluation.

[code]

Water height of the shallow-water model over time, one line per time step, with rain marked in blue.

Uncertainty quantification for mass-conserving data assimilation

LMU Munich, Chair of AI and Machine Learning, 2025

Re-implements a CNN that corrects cheap EnKF updates towards the mass-conserving QPEns solution in a modified shallow-water model, then adds uncertainty estimates on top: conformal prediction, conformalised quantile regression, evidential regression and MC dropout.

[code][write-up]

A Hopkins155 street scene with point trajectories coloured by motion: two cars and the background.

Learned trajectory embeddings for motion segmentation

LMU Munich, B.Sc. thesis, 2025

Re-implementation of Lochman et al. (CVPR 2024) for subspace clustering of point trajectories. The thesis shows that the usual Hopkins155 splits leak data, proposes a leakage-free protocol, and adds cross-trajectory attention, which lowers clustering error significantly.

[thesis][code][write-up]

3. Education and positions

Table 1. Degrees and research positions, most recent first.

Period	Degree / role	Institution
2025 – now	M.Sc. InformaticsTechnical University of MunichFocus on machine learning and 3D computer vision	Technical University of Munich
2022 – 2025	B.Sc. Computer Science, minor in StatisticsLMU MunichThesis: Learned Trajectory Embedding for Subspace Clustering (1.0)	LMU Munich
2026 – now	Research intern (remote, with the TUM Data Innovation Lab)University of Cambridge, CV4DT group	University of Cambridge, CV4DT group
2025 – 2026	Student research assistantLMU Munich, Chair of AI and Machine Learning	LMU Munich, Chair of AI and Machine Learning
2024	ScholarshipDeutschlandstipendium	Deutschlandstipendium