Globoko učenje v računalniškem vidu

Raziskovalno področje računalniški vid naslavlja probleme povezane z zajemanjem, procesiranjem, analizo in razumevanjem vizualne informacije kot so slike, video in oblaki 3D točk. Eden temeljnih problemov računalniškega vida je vizualno učenje in razpoznavanje, t.j., učenje predstavitev (predmetov, obrazov, prostorov, akcij, itn.), ki se lahko nato uporabljajo za klasifikacijo nepoznanih primerov na novih slikah. Konvolucijske nevronske mreže in sorodni pristopi globokega učenja so se izkazali za zelo uspešen način za iskanje zmogljive predstavitve in učinkovitega klasifikatorja v poenotenem pristopu, ki daje odlične rezultate v različnih nalogah s področja računalniškega vida. Glavni cilj tega predmeta je uvesti študente na področje globokega učenja, s posebnim poudarkom na aplikaciji le-tega v računalniškem vidu. Študentje se bodo seznanili z glavnimi principi računalniškega vida in strojnega učenja ter s povezavo z globokimi nevronskimi mrežami. Naučili se bodo kako naučiti in uporabljati globoke nevronske mreže z glavnim poudarkom na konvolucijskih nevronskih mrežah. Pokazali bomo kako se ti pristopi uporabljajo za klasifikacijo in detekcijo predmetov, kot tudi za ostale naloge na področju računalniškega vida in tudi širše.


The research field of computer vision addresses the problems related to acquiring, processing, analyzing, and understanding visual information such as images, videos and 3D point clouds. One of the core problems in computer vision is visual learning and recognition; i.e., learning the representations (of objects, faces, rooms, actions, etc.) that are later on used to classify unknown instances that appear in new images. This problem has been tackled since the beginning of the computer vision, however no previously proposed method has increased the performance beyond the current state of the art in such a way as deep learning approaches in the recent years. Convolutional neural networks and related deep learning approaches have proven to be a very efficient way of finding the representations and building a classifier in a unified framework that yields excellent results in various computer vision tasks. The main goal of this course is to introduce the students into the field of deep learning, with the special emphasis on its application in computer vision. The students will be acquainted with the main principles of computer vision and machine learning, relating them to neural network methods and showing them how to train and use neural networks with the main emphasis on Convolutional Neural Networks. It will be shown how these approaches can be used for object classification, localization and detection, as well as for other tasks in computer vision and beyond.