Samonadzirano učenje stereoskopske rekonstrukcije

master's thesis

• Title (english): Self-supervised learning of stereoscopic reconstruction
• Author: Jelena Bratulić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Vladimir Čeperić (član povjerenstva)
• Committee member: Željka Mihajlović (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2023-07-13, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Stereoskopska rekonstrukcija je važno i atraktivno područje računalnog vida čije primjene pronalazimo u različitim industrijama, robotici, medicini i autonomnim vozilima. Iako nadzirane metode bazirane na dubokom učenju postižu odlične rezultate, često je vrlo teško i skupo prikupiti stvarne oznake dispariteta ili dubine pa se u posljednje vrijeme naglasak stavlja na samonadzirane metode. U okviru ovog rada, proučili smo nekoliko popularnih samonadziranih metoda za stereoskopsku rekonstrukciju te smo predstavili novu metodu koju nazivamo iterativno pseudolabeliranje ansamabliranjem (IPA). U radu smo detaljno opisali našu metodu, postupke učenja i validiranja hiperparametara. Detaljni eksperimenti koristeći model RAFT Stereo pokazuju iznimno kompetitivne rezultate na skupu podataka KITTI. U radu smo naveli i moguće pravce budućeg istraživanja usmjerene u poboljšanje metode ansambliranja te uključivanje znanja semantičke segmentacije.

Abstract (english)

Stereoscopic reconstruction is a significant and appealing field in computer vision, with wide-ranging applications in industries such as robotics, medicine, and autonomous vehicles. While supervised approaches leveraging deep learning have demonstrated remarkable outcomes, acquiring disparity or depth labels often poses considerable challenges and expenses. As a result, self-supervised methods have recently garnered significant attention. In this Master Thesis, we thoroughly examine several popular self-supervised techniques for stereoscopic reconstruction and introduce a novel approach termed iterative pseudolabeling by ensambling (IPE). Our research delves into the intricate workings of IPE, including the methodology, the process for learning, and the validation of hyperparameters. Through comprehensive experimentation utilizing the RAFT Stereo model, our proposed method showcases exceptionally competitive results on the KITTI dataset. We also highlight potential avenues for future research, focusing on enhancing the ensemble method and incorporating semantic segmentation knowledge. These directions offer opportunities to further advance the effectiveness and applicability of the proposed approach.

• Keywords: stereoskopska rekonstrukcija; duboko učenje; samonadzirano učenje; računalni vid; RAFT-Stereo; pseudolabeliranje; ansambliranje; KITTI
• Keywords (english): stereoscopic reconstruction; deep learning; self-supervised learning; computer vision; RAFT-Stereo; pseudolabeling; ansambling; KITTI
• Language: croatian
• URN:NBN: urn:nbn:hr:168:710315
• 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: 2023-10-11 08:32:04