Florian Langer
Machine learning PhD student in Cambridge.
I am a machine learning PhD student at the University of Cambridge. I work in the Computer Vision group under the supervision of Dr. Ignas Budvytis and guidance of Prof. Roberto Cipolla. Specifically I have been working on 3D shape prediction from single images. More broadly I am interested in learning precise predictive models of the world from visual data. Such models include the 3D shape of objects, their texture, their material, physical properties as well as higher level information commonly associated with objects. Learning a precise model of the world is key for higher level tasks such as planning and reasoning which we commonly associate with general intelligence. Recently I have also become more interested in Neuroscience as I believe that the human brain still has many insights to offer that can radically improve artificial intelligence.
Web of Knowledge
I have started to document my process of learning more rigorously and am sharing my notes and thoughts on various books, papers and podcasts from the field of neuroscience and machine learning. In one of those books I have learned that spiders locate their prey by sensing tiny vibrations in their web with their feet. After eating the caught insect spiders will make their web harder wherever the insect was caught which will make it easier in the future to sense vibrations from this particular area of the web. In this way spiders modify their environment to store knowledge. Pretty cool!
In tribute to all the spiders out there I am calling my note section my personal Web of Knowledge.
Publications
SPARC: Sparse Render-and-Compare for CAD model alignment in a single RGB Image
F. Langer, G. Bae, I. Budvytis, R. Cipolla
33rd British Machine Vision Conference (2022)
Render-and-Compare allows for precise CAD model alignments. However, traditionally it is very slow and requires a good initialisation. It is slow because
it requires to render full objects.
it then requires to process fully-rendered images.
a similarity function between the image and the render is maximised via gradient descent which requires a large number of iterations (100s to 1000).
It requires a good initialisation because
4. maximising the similarity function may get stuck in local optima.
In this work we address these short comings. We address 1. and 2. by only sampling a set of sparse points and surface normals from the CAD model to be rendered and then processing only those sparse inputs as opposed to a full image. We address 3. and 4. by using a network to directly predict pose updates rather than a similarity function which reduces the number of iterations needed to just 3 and simultaneously makes the network robust to object initialisations.
Leveraging Geometry for Shape Estimation from a Single RGB Image
Predicting 3D shapes and poses of static objects from a single RGB image is an important research area in modern computer vision. Its applications range from augmented reality to robotics and digital content creation. In this work we demonstrate how cross-domain keypoint matches from an RGB image to a rendered CAD model allow for more precise object pose predictions compared to ones obtained through direct predictions. We further show that keypoint matches can not only be used to estimate the pose of an object, but also to modify the shape of the object itself. This is important as the accuracy that can be achieved with object retrieval alone is inherently limited to the available CAD models. Allowing shape adaptation bridges the gap between the retrieved CAD model and the observed shape.
Expansion by regions with pySecDec
G. Heinrich, S. Jahn, S.P. Jones, M. Kerner, F. Langer, V. Magerya, A. Poldaru, J. Schlenk, E. Villa
Computer Physics Communications (2021)
We discuss the technique of expansion by regions from a geometric perspec- tive, and its implementation within pySecDec, a toolbox for the evaluation of dimensionally regulated parameter integrals. The program offers an auto- mated way to perform asymptotic expansions and provides a new mechanism for efficiently evaluating amplitudes, as well as individual integrals. The usage of the new features available within pySecDec is illustrated with several examples.
Shape and Pose Estimation of Rigid Objects from a Single RGB Image
This work addresses the problem of estimating 3D shapes and poses of rigid objects from a single RGB image. With applications ranging from augmented reality to robotics and digital content creation this task is getting increasingly more attention in the research community
Learning on the Fly - Precise Positioning of a Drone using Spoken Language Commands
We build a system that takes voice commands and navigates to a certain location.
Bosonic String Theory and T-duality
As string theory may serve as both, a theory of quantum gravity and a unified theory of all fundamental forces, it has naturally been one of the most active research fields of theoretical physics in the last decades. Known for its notorious difficulty, this report aims to be a simple introduction to bosonic string theory and one of its beautiful symmetries, T-duality.
Starting off with the general idea and a brief historic background, the classical string is analysed. The Polyakov action is investigated in terms of its equations of motion and symmetries. These symmetries are used to fix the flat gauge in which the equations of motion take a particularly simple form. Following a naive quantisation procedure the constraints arising from the equations of motion are imposed as operator conditions on the Hilbert space. This Hilbert space contains negative norm states, but using the residual conformal symmetry it is shown how these can be removed. In a second method of quantisation, the BRST procedure, the original symmetry is replaced by a larger symmetry that is still present after gauge fixing. It is demonstrated that physical states appear as the cohomology of the generator of the BRST symmetry. Finally the concept of compactification is introduced and it is shown how string theories with differing compactified dimensions are equivalent under T-duality.
Education
PhD in the Computer Vision Group - University of Cambridge
I am doing a PhD in the Computer Vision group under the supervision of Dr. Ignas Budvytis and guidance of Prof. Roberto Cipolla. I am working on 3D shape prediction from single images. More broadly I am interested in learning precise predictive models of the world from visual data.
Since 2020
MPhil in Machine Learning and Machine Intelligence - University of Cambridge
I completed the MLMI MPhil with a distinction. My master thesis "Learning on the Fly - Precise Positioning of a Drone using Spoken Language" was supervised by Dr. Ignas Budvytis.
2019 - 2020
MSci & ARCS in Phyiscs with Theoretical Physics -
Imperial College London
Dean’s list award 2017/2018 and 2018/2019 for top 10% of students in each year cohort.
My Master thesis “Bosonic String Theory and T-duality“ was supervised by Prof. Daniel Waldram.
2015 – 2019
Abitur
Salem International College
2012 – 2015
I completed the German Abitur with the top grade 1.0 (865 out of 900 points).
Awards: National maths competition “Bundeswettbewerb für Mathematik” 2015 Third Prize, 2014 Official Recognition, “Mathematikolympiade” Federal state final Hessen 2011 7. Place, Hessen 2009 1. Place
Work Experience
Research Internship - Max-Planck Institute for Physics
Summer 2018
I completed a two months summer research project in the phenomenology group. As part of the project I implemented aspects of the expansion by regions methods that allows calculating higher order corrections to loop integrals for particle interactions in Quantum Field Theory.
Founder of Emote Reality - Virtual Reality Start-Up
In Emote Reality we specialised in creating virtual tours for businesses that allow their customers to experience their locations through Virtual Reality glasses or any smart device.
2016 -2018
Internship at Staramba SE
Summer 2017
Staramba SE started by creating virtual scans of people which they then proceeded to animate for interactive VR experiences. I was working with the level artists in high-poly modelling for constructing an extremely realistic virtual world. They are now called NEXR.