Luca Lanzendörfer

lanzendoerfer at ethz.ch

PhD student at ETH Zurich

I’m currently working as a PhD student at ETH Zurich, advised by Prof. Roger Wattenhofer. My work focuses mainly on applications of Deep Learning. Previously, I worked as a Machine Learning Engineer on Natural Language Processing, Computer Vision and Recommender System solutions in Zurich and on self-driving cars in Silicon Valley. I received my graduate degree at ETH Zurich in Computer Science focused on Visual Computing and Machine Learning.

Previous Projects

Project for the Deep Learning course of 2017 at ETH Zurich.

Abstract- We propose a model for Visual Question Answering (VQA) based on iBOWIMG. Our model uses image features with attention extracted from InceptionV3 as well as object features extracted from the VQA dataset using YOLO object detection. Our model is able to achieve a competitive score on the VQA v1 test-dev scoreboard. Furthermore, we analyse shortcomings of the dataset and explore the current state of VQA.

Project for the Computational Intelligence Lab of 2017 at ETH Zurich.

Abstract- Collaborative Filtering methods have become widely used in consumer oriented e-commerce through different matrix factorization methods. We demonstrate how Collaborative Filtering methods based on matrix factorization can be further improved by boosting many specialized SVD and Neural Network approaches to obtain a competitive score on the CIL dataset.

Project for the Game Programming Lab of 2017 at ETH Zurich.

Abstract- A 2D top-down split screen multiplayer racing game. The game features pixel art graphics and contains various features that make multiplayer fun and engaging. Additionally, to further increase replayability we designed a formal grammar to procedurally generate race tracks. Programmed using MonoGame Engine in C# to run on Windows PC and Microsoft Xbox.

Project for the Physically-based Simulation course of 2016 at ETH Zurich.

Abstract- Simulation of a photorealistic fountain water show using a variation of Smoothed Particle Hydrodynamics called PCISPH. We created a custom implementation of Marching Cubes for the surface reconstruction. Visualization was done using the physically-based Mitsuba renderer. We simulated 1 mio. particles for around 100’000 timesteps on the ETH Euler compute cluster. Each reconstructed surface contains 4 mio. vertices and 10 mio. faces and we ray-traced around 250 mio. rays per frame.