Johan Edstedt

I am a master student at Linköping University, studying Signal & Image Processing. I also recently finished an exchange year at Katholieke Universiteit Leuven where I focused on learning more about A.I. My main interests are in generative modelling and image processing. I'm currently doing my Master Thesis, expected to be done by June 2020 + Working for FIA Robotics, implementing Computer Vision Solutions for the @home-cup 2020.

Email  /  Github  /  LinkedIn

profile photo
Education
Boundary_png

Linköping University
Master of Science
Applied Physics and Electrical Engineering
Specialization in Image and Signal Processing

Aug 2014 - June 2020

Boundary_png

Katholieke Universiteit Leuven
Erasmus Studies
Focus on Artificial Intelligence

Sep 2018 - June 2019

Projects

Below are some sample projects I've worked on.

CDIO Project - Automated Analysis of Cells

We worked together with TB researchers at LiU to develop new methods to analyze the interaction between Macrophages and Tuberculosis

I worked on implementing the tracker, mainly working on semantic segmentation + fusing motion estimates. I was also partly responsible for data management, data collection, and version control.

Bachelors Project - Autonomous Driving (Report in Swedish)

The project involved constructing an autonomous race car, navigating within the bounds of a walled off track. We used ultrasonic sensors and a LIDAR as our main source of control. We combined control theory with an ANN to efficiently control the vehicle.

I worked on integrating the LIDAR and ultrasonic sensors into the system, as well as data collection for ML and dependency management.

Exploring Multi-Agent Learning - Opponent modelling & Deep reinforcement learning

We investigated agent behaviour in a Multi-Agent system. We were given a simulated environment and constructed Deep Learning RL-agents with different kinds of reward functions, and investigated their observed behaviour.

I mainly worked on the models, i.e. implementing the network architecture as well as the RL algorithm (A3C).

Visual Question Answering

The main purpose was to construct a multi-modal question answering system that could answer text based questions about a specified image.

I was responsible for both the architecture and training of the model. We improved upon the results of the original paper, but we did not achieve state of the art performance.

Efficient construction of random forests

The goal was to build an unpruned random forest using gini impurity as split heuristic as efficiently as possible using C++. Implementation is approx. 3x faster than scikit-learn.


Website inspired from here.