Main | Photography | Etc.

Vít Růžička

Welcome to my webpage, here are few details about me, my projects in academia and photography. Currently, I am a DPhil student at the Department of Computer Science at the University of Oxford, supervised by Niki Trigoni and Andrew Markham. In 2019-20, I was working as a research assistant and lecturer at the University of the Arts London in the Creative Computing Institute. Before that I was on a research internship at ETH Zurich in the EcoVision group (2019) and at Carnegie Mellon University in Franz Franchetti’s group (2017-18). I did my MSc and BSc at the Czech Technical University in Prague. Recently, I have been working with Frontier Development Lab, as a researcher in FDL Europe 2021 and as a Machine Learning team lead in FDL USA 2022. In the summer of 2023, I also did an internship at the ESA Φ-lab. I like Machine Learning research, Arts, literature, traveling and analog photography.

Email  /  Google Scholar  /  GitHub  /  YouTube  / 

My research interests are in Machine Learning applied in image analysis and generation, Computer Vision and Generative Models, often in the area of Remote Sensing. For the full list of my publications please check my Google Scholar profile, few selected works are listed here.

We introduce a new dataset of hyperspectral images from the AVIRIS-NG sensor with refined methane labels. We propose lightweight machine learning models for semantic segmentation of methane plumes, which outperform the classical baselines. We show zero-shot generalisation in the new EMIT hyperspectral sensor. Work conducted with Trillium Tech: STARCOP.

We show how to apply the contrastive learning technique of SimCLR for the task of unsupervised change detection, exemplified on wildfire detection. For this project, I was the ML team lead at the FDL US 2022.

Presentation at AI+HADR'21:
 

We introduce the RaVÆn system, a lightweight, unsupervised approach for change detection in satellite data based on Variational Auto-Encoders (VAEs) with the specific purpose of on-board deployment. It flags changed areas to prioritise for downlink, shortening the response time. We show that the proposed method outperforms pixel-wise baselines and we test it on resource-limited hardware. We also release the annotated dataset of extreme events. Work conducted at the FDL 2021. In 2023 our model was tested on-board the D-Orbit's ION satellite which we report in the paper published at IGARSS 2023 - in addition, we showcase training of Machine Learning models on-board of the satellites (read more about it here).

Example of inference:
 
Figure shows aerial images from years 2012 (left), 2015 (middle) and the predicted change label (right). 

Presentation at MACLEAN'20:
 

We investigate active learning in the context of deep neural networks for change detection and map updating. For this task we propose a Siamese version of the U-Net model and test two approaches in model uncertainty prediction (model ensembles and Monte Carlo Batch Normalization). Iteratively selecting the most uncertain samples for annotation, we reach the same performance as a model supervised with a large fully annotated training set with approx. 99% fewer annotated samples.

Presentation at HPEC'18:
 

Paper in which we scale up the YOLO v2 object detection model to process video of 4K-8K resolution in 3-6 fps using GPUs. We suggest a two-stage processing pipeline which allows us to ignore some areas of the image. This is much faster than processing the whole image yet retains 98% of the accuracy on a real-world video footage dataset.

I always found art to be a way of exploring and communicating ideas, getting to understand or describe the unknown. In a way scientific work has similar goals but uses different language and methods. On this track I am co-organizing the yearly ArtSci exhibition at ETH.

My own creative practice started with analog photography (manual and slow process of taking and developing pictures in a darkroom) and later got mixed in with Machine Learning. Can an automatic programmed or trained process be creative on its own? Perhaps not, but with a bit of work and understanding, it can be used as a tool of creative expression. Great examples of this can be seen in the neurography of Mario KlingemannMario Klingemann - Memories of Passersby I, 2018:
 
Source: aiartists.org/mario-klingemann, dataset handling of Anna RidlerAnna Ridler - Myriad (Tulips), 2018:
 
 , domain transfers of Memo AktenMemo Akten - Learning to See (Gloomy Sunday), 2017:
 
Image reconstructed using a model trained on particular visual dataset. or lectures by Gene Kogan. If you would like to know more, you can hear me introducing this area in a video recordingSlides from the first lecture:
 
Introductory session exploring the intersections of Art and Machine Learning.  of my first class of Exploring Machine Intelligence at UAL.

A short experimental film made about AI with AI. It was premiered at the Ji.hlava IDFF 2023 with this synopsis: "24 hours, 39 minutes and 35 seconds. One Sol, during which the author lets us empathically experience the inner and outer journey of an aging rover on Mars, whose view of the great unknown tells us much more about ourselves than any scientific research."

This project explores what happens when we use super resolution models to enhance our data. Recorded reality is more and more encoded and mixed with the artefacts of the neural age. This offers possibilities for creative expression, as long as these enhancements are in our hands. We ask alongside with "The end" screens: imagine the possible futures, what happens next?

This is a short experimental film which explores the latent space of generative models trained on my own collected and curated selections of scenes. The concept of film editing is replaced by escalating jumps in the latent space, montage of feature vectors and sequences based on similarity rather than time. Imagery is reinterpreted and reordered following a compressed form.

A neuro-photographic collage playing with the typography of Amazonian jungle patterns, shapes and textures and their reinterpretation using generative neural networks and the material properties of Super8 analog film. Digital generative models as a extension of a limited number of film frames. Analog film footage as a mode of interaction with the medium inside a darkroom with photographic paper and scissors.

In this project we generate endless non-existent variation of 3D patches of landscape with deep learning models trained on the features of real-world landscape. The name of the project is based on the works of Ernst Jentsch "On the Psychology of the Uncanny" and Masahiro Mori "The Uncanny Valley", which characterize the "uncanny" as a mental state which occurs when one is not able to distinguish between the imagined and the real or between the dead and the living. Generated samples resemble the real, but are in fact a reimagination of the original dataset.

Presentation at ISEA'19:
 

We use LSTM models to generate fake newspaper articles focusing on areas often targetted as the contents of fake news (topics such as politics or fake news themselves). While the purpose of most fake news is misinformation and political propaganda, we explore it as a new type of myth created in the age of internet identities. These generated articles serve as a mirror and can give us insight into people’s hidden fear and desire similir to myths, folk tales or urban legends.

Ox, CDT AIMS - DPhil course, practicals

Ox, CS - Professional Masters course, labs

UAL, CCI - MSc course "IU000133 Artificial Intelligence for the Media"

UAL, CCI - MSc course "IU000128 Coding Three: Exploring Machine Intelligence"

UAL, CCI - BSc Year 1 course "Data, Math and Methods"

Webpage design from here.