Unit:
Κατεύθυνση Μεγάλα Δεδομένα και Τεχνητή ΝοημοσύνηΠληροφορική
Author:
Efthymiadis Nikolaos
Supervisors info:
Αβρίθης Γιαννης, Research Scientist, Inria Rennes-Bretagne Atlantique, France
Εμίρης Ιωάννης, Καθηγητής, Τμήμα Πληροφορικής και Τηλεπικοινωνιών, ΕΚΠΑ
Τόλιας Γιώργος, Επίκουρος Καθηγητης, Department of Cybernetics, Czech Technical University in Prague
Original Title:
ZERO-SHOT DOMAIN ADAPTATION FOR SKETCH RECOGNITION
Translated title:
ZERO-SHOT DOMAIN ADAPTATION FOR SKETCH RECOGNITION
Summary:
In this thesis we are tackling the zero-shot domain adaptation problem between the image and the sketch domains. By "zero-shot" we mean in term of sketches. The term "domain adaptation" suggests that we are going to learn with images as examples and we will use that knowledge to predict sketches.
In order to do this we are going to make one transformation to the image realizations and one transformation to the sketch realizations such that the transformed realizations can be seen as they were sampled from the same random variable. After that our strategy is to learn a function on the transformed realizations of the image domain and predict our sketches normally.
The transformations we are going to use have as an objective to represent both domains as one-pixel thick binary edge maps, encoding with 1 the existence of an edge and with 0 the absence of it. For the image domain a combination of edge detection, thresholding and thinning procedures is going to be used in a data augmentation philosophy. For the sketch domain we will use a deterministic thresholding rule followed by a thinning procedure.
Main subject category:
Technology - Computer science
Keywords:
Zero-Shot Learning, Domain Adaptation, Sketch Recognition
