I obtained my bachelor and master's degree in Geomatics Engineering from ETH Zürich. During my master's
specialized in deep learning, computer vision and remote sensing.
I am expecting to graduate in Summer/Autumn 2025. I am looking for an industry position.
Research
I'm interested in computer vision, deep learning, and their applications to remote sensing.
Most of my work is related to super-resolution, depth estimation or both at the same time.
Some papers are highlighted. * indicates equal contribution.
Marigold is an affine-invariant monocular depth estimation method based on Stable Diffusion,
leveraging its rich prior knowledge for better generalization and achieving state-of-the-art performance with significant improvements,
even with synthetic training data.
POPCORN is a lightweight population mapping method using free satellite images and minimal data,
surpassing existing accuracy and providing interpretable maps for mapping populations in data-scarce regions.
We present a novel approach for arbitrary-scale single image super-resolution (ASSR) that uses neural fields with an adaptive Gaussian point spread function (PSF) to prevent aliasing and achieve superior results,
offering more parameter efficiency and setting a new state of the art while maintaining computational efficiency.
We propose DADA, a novel approach to depth image super-resolution by combining guided anisotropic diffusion with a deep convolutional network, enhancing both edge detail and contextual reasoning.
This method achieves unprecedented results in three benchmarks, especially at larger scales like x32
Information about the focal length with which a photo is taken might be obstructed (internet photos) or not available (vintage photos).
Inferring the focal length of a photo solely from a monocular view is an ill-posed task that requires knowledge about the scale of objects and their distance to the camera - e.g. scene understanding.
I trained a deep learning model to acquire such scene understanding to predict the focal length and open-source the model with this repository.
POMELO is a deep learning model that creates fine-grained population maps using coarse census counts and open geodata,
achieving high accuracy in sub-Saharan Africa and effectively estimating population numbers even without any census data.
We propose a method for forecasting the emergence and timing of new buildings using a deep neural network with a custom pretraining procedure, validated on the SpaceNet7 dataset.
We propose using neural ordinary differential equations (NODEs) combined with RNNs to improve crop classification from irregularly spaced satellite images,
showing enhanced accuracy over common methods, especially with few observations,
and better early-season forecasting due to the continuous representation of latent dynamics.
This work presents a method for automatically refining 3D city models generated from aerial images by using a neural network trained with reference data and a loss function to improve DSMs,
effectively preserving geometric structures while removing noise and artifacts.
News
16th - 21st of June 2024: I'm attending CVPR 2024 in Seattle, WA, USA presenting Marigold.
6th June 2024: Invited talk for the KoRaTo Team of University of Ulm at ETH Zurich.
25th April 2024: Invited talk about my population mapping projects at WorldPop in Southampton, UK.
December 2023 - February 2024: I'm interning at Meta's Reality Labs in Redmond, WA, USA.
22nd of September 2023: Invited talk at UZH Astrophysics Seminar. Recording