VQVAE

Vector Quantized Variational Autoencoders are a state of the art method to generate high fidelity images. To gain a better understanding I reimplemented them in JAX and trained them on topographic maps from swisstopo, to generate an endless supply of new places to explore.

This is work in progress, code and demo to be released soon.

Learning 3D Point Cloud Descriptors for Robust Localization

In this project I worked on improving localization by learning descriptors robust to changes in viewpoint, illumination and deformation. The main idea was to incorporate geometric and semantic information, using 3D point clouds instead of typical 2D images.

This work was published at IROS 2021 and can be found on arXiv.

More on this project here.

Deep Learning-based Plastic Detection on Beaches using MAVs

As part of The Plastic Tide effort I designed and built a machine learning system to survey beaches for plastic waste, detect and characterize the waste and estimate its volume.

More on this project here.

MapLab

While working at the Autonomous Systems Lab I contributed to one of the largest open source SLAM frameworks, Maplab. Among the things I worked on were adding wheel constraints to the optimization to reduce trajectory drift and support for the semantic point cloud descriptors I developed.

Collaborative Filtering: Stacking Collaborative Filtering and Neural Networks for Improved Recommendations

This project was done together with Kevin Klein and Lorenz Kuhn.

Abstract: Online businesses face the challenge of recommending relevant products to users based on users’ previous preferences and similar customers. This work explores the use of classic matrix factorization methods on the one hand and recent neural network-based methods on the other hand. Final predictions were further improved using ensembling methods such as bagging and stacking. We report similar, competitive scores for matrix factorization methods and slightly lower accuracy for neural network-based methods with a final ensemble RMSE of 0.964.

The full report can be found here.