Skip to content
Leonardo Vanni

Leonardo Vanni

3D/4D computer vision · BSc in Mathematics & CS for AI, Bocconi University

Real scenes move, break, and fall apart. I build 3D vision that keeps up — and knows when it's wrong.

Currently: topology-aware 4D Gaussian Splatting for egocentric video (TopoEgo), and SE(3)-consistent uncertainty for 3D foundation models (thesis).

Currently looking for research internships in 3D computer vision.

Portrait of Leonardo Vanni

Selected research

TopoEgo ongoing Jan 2026 – present

Objects get cut. Most 4D methods pretend they don't.

A piece-wise rigid kinematic tracker for 4D Gaussian Splatting: at topological events, canonical Gaussians split into independent SE(3) branches instead of being smeared by a continuous deformation field. Comes with a benchmark of cut/open/break interactions curated from EPIC-KITCHENS and Ego4D.

First author — with Chiara Plizzari (Bocconi University), Stefano Gasperini (TUM)

Paper, code, and the TopoEgo benchmark will be released publicly upon acceptance.

SE(3) Uncertainty for VGGT completed 2026

A 3D foundation model that knows when it's wrong.

Extends VGGT with a decoupled branch predicting full 6×6 camera-pose covariances on the SE(3) manifold. The uncertainty calibrates with a single temperature, structurally matches Bundle-Adjustment covariances, and spikes exactly on mislocalized frames — a built-in failure detector.

Sole author (BSc thesis) — supervised by Prof. Alessandro Pigati, Bocconi University

Full research page →

What I care about

Topology is not noise

When objects split, the representation should split too — not stretch, ghost, or hallucinate melted cheese.

Know when you're wrong

A pose without a covariance is an opinion. Calibrated uncertainty is what lets 3D models plug into fusion, planning, and robots.

Geometry as inductive bias

Lie groups and manifold-aware losses aren't formalism for its own sake — they're how models stay correct under the transformations the world actually applies.

About

I’m completing my BSc in Mathematics and Computing Sciences for AI at Bocconi University (graduating July 2026), with an exchange semester at École Polytechnique. My current research is TopoEgo, a first-author project on topology-aware 4D Gaussian Splatting for egocentric video, with collaborators at Bocconi and TUM. My bachelor thesis extended the VGGT 3D foundation model with SE(3)-consistent uncertainty estimation. Before that: computer-vision lead on an EEG-controlled prosthetic arm, medical computer vision at Vision Dental, ML engineering at Accenture.

News