Research
I'm broadly interested in computer vision and learning. I develop models that extract high-level information of the world to assist robots and automated systems such as self-driving cars. Representative publications are highlighted.
|
|
Vulcan Centaur: towards end-to-end real-time perception in lunar rovers.
De Curtò and
Duvall.
We introduce a new real-time pipeline for SLAM and VIO in the context of planetary rovers. We leverage prior information of the location of the lander to propose an object-level SLAM approach that optimizes pose and shape of the lander together with camera trajectories of the rover. As a further refinement step, we propose to use techniques of interpolation between adjacent temporal samples; videlicet synthesizing non-existing images to improve the overall accuracy of the system.
|
|
Cycle-consistent Generative Adversarial Networks for Neural Style Transfer using data from Chang’E-4.
De Curtò and
Duvall.
dataset
We introduce tools to handle planetary data from the mission Chang’E-4 and present a framework for Neural Style Transfer using Cycle-consistency from rendered images. The experiments are conducted in the context of the Iris Lunar Rover, a nano-rover that will be deployed in lunar terrain in 2021 as the flagship of Carnegie Mellon, being the first unmanned rover of America to be on the Moon.
|
|
Doctor of Crosswise: Reducing Over-parametrization in Neural Networks.
Curtò,
Zarzà,
Kitani,
King and
Lyu.
code
Dr. of Crosswise proposes a new architecture to reduce over-parametrization in Neural Networks. It introduces an operand for rapid computation in the framework of Deep Learning that leverages learned weights.
|
|
High-resolution Deep Convolutional Generative Adversarial Networks.
Curtò,
Zarzà,
Torre,
King and
Lyu.
dataset /
supplement /
video
In order to boost network convergence of DCGAN and achieve good-looking high-resolution results we propose a new layered network, HDCGAN, that incorporates current state-of-the-art techniques for this effect.
Recognitions:
State-of-the-art in synthetic image generation on CelebA 128x128 (MS-SSIM). 2017.
State-of-the-art in synthetic image generation on CelebA 64x64 (FID). 2017.
|
|
Segmentation of Objects by Hashing.
Curtò,
Zarzà,
Smola and
Gool.
We propose a novel approach to address the problem of Simultaneous Detection and Segmentation. We use an efficient and accurate procedure that exploits the feature information of the hierarchy using Locality Sensitive Hashing.
|
|
McKernel: A Library for Approximate Kernel Expansions in Log-linear Time.
Curtò,
Zarzà,
Yang,
Smola,
Torre,
Ngo and
Gool.
code /
slides /
coverage
McKernel introduces a framework to use kernel approximates in the mini-batch setting with Stochastic Gradient
Descent (SGD) as an alternative to Deep Learning.
|
Miscellaneous
From Catalunya (Regne d'Espanya); in Pittsburgh, Hong Kong and Zürich. My last name is 'De Curtò', 'DíAz' is the surname of my mother; 'i/y' is a conjunction that literally means 'and' in Catalan/Spanish. 'De Curtò' signifies 'to cut' in Latin.
A tribute to some of my family, ancestors and mentors: De Curtó i Bel (my father): passionate about cars, mechanics and painting. De Curtó i Berengué (my grandpa): he survived the civil war and protected the family until just after I was born. Bel i Bosch (my grandma): she knew beforehand I was the chosen one. Supported my education and all my crazy endeavors without asking for an explanation, she truly believed in me. Masterson (my teacher of English from age 7 until 17, and onwards a real supporter and a second mother): from Edinburgh (Scotland). She has taught me how to be a good Briton.
|
|