Invertible Conditional GANs for image editing
Guim Perarnau, Joost van de Weijer, Bogdan Raducanu and Jose M. Álvarez
November 2016
Samples generated by the IcGAN This work was originally conceived as my master thesis. It was later accepted and selected for an oral presentation at the NIPS 2016 Workshop on Adversarial Training.

What if you could take a picture of your face and have a neural network arbitrarily change how you look? This is exactly what the IcGAN does using the power of Generative Adversarial Networks. It doesn't apply to faces only, but to whatever image dataset you train with.

Video surveillance for road traffic monitoring
February 2016

My teammates and I analyze different approaches to effectively monitor traffic. Considering the limitations of the real world (e.g. camera jittering, dynamic background), we use computer vision techniques to track vehicles and estimate their speed with an RGB camera.

Traffic sign detection and recognition
February 2016

We built a detector and classifier able to recognize 15 different types of traffic signs with 81.05% precision and 75.68% F1-Score. We experimented with hand-crafted features and neural networks.

Map generation with a UAV
December 2015
[Article] (in Catalan)
Example of the map generation process
Awarded as the best final degree project.

Given an aerial video of a terrain captured with a UAV equipped with a sensor, I learn and explore different techniques to accurately represent the map where the UAV has been.