Usporedba dubokih modela i tradicionalnih metoda za 3D registraciju oblaka točaka

master's thesis

• Title (english): Deep Learning Vs. Traditional 3D Point Cloud Registration Methods
• Author: Fran Čudina
• Mentor: Tomislav Pribanić (mentor)
• Committee member: Tomislav Pribanić (predsjednik povjerenstva)
• Committee member: Mario Cifrek (član povjerenstva)
• Committee member: Tomislav Petković (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-08, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Kao jedan od bitnih iskoraka u računalnom vidu, fokusiranom na 3D oblake točaka, je i 3D registracija koja se svodi na izdvajanje značajki iz točaka na danom oblaku točaka i provođenju spajanja dva oblaka točaka prema podudaranjima između njih. Ovakav pristup se u većini slučajeva preuzima iz 2D domene i prilagođava na 3D točke stoga je takav proces prilagodbe i opisan. Primarno u ovom radu analiziramo tradicionalne metode i metode bazirane na dubokim modelima i korištenu teorijsku i matematičku pozadinu svake od metode zasebno. Od tradicionalnih metoda obrađena je Harris3D metoda s FPFH metodom, te Super4PCS metoda i njena sporija verzija 4PCS, dok od dubokih analiziramo D3Feat i 3DMatch metodu kao jedne od često korištenih metoda prilikom usporedbe rezultata dubokih modela. Provedena je i evaluacija svih metoda na 3DMatch skupu podataka gdje su duboke metode nadmašile tradicionalne za veliku razliku te su dobiveni rezultati prikazani tablično i kroz provedene primjere 3D registracije.

Abstract (english)

As one of the important steps in computer vision, focused on 3D point clouds, is 3D registration, which comes down to extracting features from points on a given point cloud and conducting the merging of two point clouds according to the matches between them. This approach is in most cases taken from the 2D domain and adapted to 3D points, so such an adaptation process is described. Primarily in this paper we analyze traditional methods and methods based on deep models and the used theoretical and mathematical background of each of the methods separately. Of the traditional methods, Harris3D method with FPFH method, Super4PCS method and its slower version 4PCS, while from deep methods we analyze D3Feat and 3DMatch method as one of the most frequently used methods when comparing the results of deep models. An evaluation of all methods on the 3DMatch data set was performed, where deep methods surpassed the traditional ones by a large margin, and the obtained results were presented in tables and through implemented examples of 3D registration.

• Keywords: Detekcija ključnih točaka; deskripcija točaka3D registracija; 3DMatch; 4PCS; Harris3D; FPFH; D3Feat
• Keywords (english): Keypoint detection; point description; 3D registration; 3DMatch; 4PCS; Harris3D; FPFH; D3Feat
• Language: croatian
• URN:NBN: urn:nbn:hr:168:321271
• Study programme: Title: Computing; Study programme type: university; Study level: graduate; Academic / professional title: magistar/magistra inženjer/inženjerka računarstva (magistar/magistra inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-14 08:39:34