Research

Most of my work sits where 3D vision meets geometry — on the messy end of it. Real video, especially egocentric video, keeps doing things current methods assume away: objects get cut, opened, broken, taken apart; models return confident answers from frames they have badly mislocalized. I care about representations that model these events explicitly, and about models that report calibrated confidence instead of a bare point estimate.

A secondary thread applies the same taste for structure to formal mathematics: discrete diffusion models for theorem proving in Lean, with the Lean kernel as an automated verifier.

TopoEgo ongoing Jan 2026 – present

TopoEgo: Topology-Aware 4D Gaussian Splatting for Egocentric Vision

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

Current 4D Gaussian Splatting methods assume scene topology never changes: a connected object must remain connected. Egocentric video violates this constantly — cutting, opening, breaking, disassembly — and continuous deformation fields respond with severe stretching and ghosting artifacts. TopoEgo extends rigid tracking into a piece-wise rigid kinematic tracker, using spacetime clustering to route canonical Gaussians into independent post-event SE(3) branches at topological boundaries. The project also introduces the TopoEgo benchmark: curated EPIC-KITCHENS and Ego4D sequences with camera poses, dense 2D segmentation masks, and temporal state-change annotations for discontinuous egocentric interactions.

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

Geometrically-Grounded Uncertainty Quantification for Foundational 3D Vision Models

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

Problem. Feed-forward 3D foundation models made reconstruction fast and robust, but they answer with a single deterministic guess. Downstream systems — sensor fusion, active vision, robotics — need to know how much to trust each pose, not just what it is.

Idea. Keep VGGT’s state-of-the-art mean prediction frozen, and train a parallel lightweight branch to predict a full $6 \times 6$ covariance over each camera pose, formulated properly on $\mathrm{SE}(3)$ : body-centric perturbations on the Lie algebra, a left-invariant metric that reconciles meters with radians, and a scale-aware negative log-likelihood with a curriculum that prevents degenerate collapse.

Why it matters. On CO3D the learned covariances calibrate with a single temperature and structurally match the analytical covariances of classical Bundle Adjustment; on EPIC-KITCHENS the predicted variance spikes exactly on mislocalized frames. That turns a black-box reconstructor into a component you can put inside a probabilistic pipeline — and gives it a failure detector for free.

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

Thesis page

Prosthetic Arm Vision public Sep 2024 – Jun 2025

Intelligent Prosthetic Arm — Real-Time Vision for Grasp Validation

Real-time vision for a hand that can't afford to guess.

Problem. An EEG-controlled prosthetic hand gets a noisy, low-bandwidth intent signal from the user. Before it closes around an object, something has to check — in real time, from a single camera — whether the grasp is actually feasible.

Idea. Fuse object detection, monocular depth, segmentation, and hand-pose tracking into Kalman-filtered 3D state for both the hand and the target, then validate the intended grasp with a hybrid model that reasons about 3D object geometry and hand pose together.

Why it matters. Grasp validation is the safety layer between “the user thought about grasping” and “the motors move.” The pipeline runs live, ships with a GUI demo and a Docker setup, and the report is public.

Computer Vision Lead / Project Lead (BAINSA × Politecnico di Milano)