Semantična segmentacija slik za razpoznavanje notranjih prostorov

Lampe, Ajda (author), Kristan, Matej (mentor)

Razpoznavanje prostorov je zanimiv problem na področju računalniškega vida, ki je praktično uporaben na mnogo področjih v vsakdanjem življenju. Z razvojem mobilne robotike bo potreba po učinkovitem in točnem razpoznavanju prostorov rasla. V zadnjem času metode za klasifikacijo prostorov dosegajo vedno boljše rezultate z uporabo konvolucijskih nevronskih mrež, naučenih na veliki količini podatkov, vendar večina metod temelji na razpoznavanju celotne slike. Slabost teh sistemov se pokaže, kadar se na sliki pojavi več kot en prostor. V diplomskem delu smo razvili metodo, ki slabost obstoječih metod rešuje s semantično segmentacijo, pri tem pa smo se osredotočili na osem najpogostejših kategorij notranjih prostorov. Z uporabo dopolnjene in predelane zbirke podatkov smo izdelali in naučili tri konvolucijske nevronske mreže, ki se med seboj razlikujejo v številu polno povezanih nivojev. Njihovo točnost segmentacije in pravilnost detekcije smo numerično ovrednotili in vrednosti primerjali z rezultati obstoječe klasifikacijske mreže, ki dosega odlične rezultate pri klasifikaciji na nivoju slike. Rezultate mrež smo analizirali tudi kvalitativno. Naučene mreže presegajo rezultate referenčne, trenutno najboljše, metode za slabih 40\% pri lokalizaciji prostora in za 20\% pri detekciji objektov v sliki.

• Language: Slovenian
• Keywords: računalniški vid, razpoznavanje prostorov, semantična segmentacija, konvolucijske nevronske mreže
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-85565
• Publication date in RUL: 16.09.2016

LAMPE, Ajda, 2016, Semantična segmentacija slik za razpoznavanje notranjih prostorov [online]. Bachelor’s thesis. [Accessed 22 March 2025]. Retrieved from: https://repozitorij.uni-lj.si/IzpisGradiva.php?lang=eng&id=85565

Secondary language

• Language: English
• Title: Semantic segmentation of images for indoor place recognition

Abstract

Space recognition is an interesting computer vision problem with many practical applications. Improvements in field of mobile robotics will most likely increase the need for efficient and accurate scene recognition systems. Lately, room classification methods have reached high classification accuracy with the use of popular convolutional neural networks, trained on large datasets, but most of the methods are based on holistic classification. Their disadvantage shows when presented with an image of multiple places. In this thesis we present a method that addresses the disadvantage of existing methods by use of semantic segmentation. In the work we focus on recognizing 8 most common indoor place categories. We improved and changed an existing dataset according to the problem and used it to build and train three convolutional neural networks with different numbers of fully-connected layers. We evaluated their segmentation and detection accuracy with use of mean intersection-over-union measure and F-measure, respectively, then compared obtained results with those of an existing holistic classification network, which achieves state-of-the-art results on the task of image-level classification. We also give a qualitative analysis of trained networks' results. Results show that our method outperforms the current state-of-the-art method by almost 40\% on the task of place localization and by 20\% on the task of place recognition.

• Keywords: computer vision, place recognition, semantic segmentation, convolutional neural networks