Sustav za segmentaciju videozapisa

master's thesis

• Title (english): System for Segmentation of Video Sequences
• Author: Fran Pugelnik
• Mentor: Tomislav Hrkać (mentor)
• Committee member: Tomislav Hrkać (predsjednik povjerenstva)
• Committee member: Siniša Šegvić (član povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-07-11, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Ovaj rad se fokusira na pronalazak načina ubrzanja postupka segmentacije fotografije kako bi se taj postupak mogao upotrijebiti za segmentaciju videozapisa. Prvo je istražen sustav R-CNN (i njegove kasnije iteracije) za detekciju objekata, te Mask R-CNN za segmentaciju instanci. Proučivši te sustave i procijenivši njihove pozitivne aspekte i nedostatke okrenuli smo se prema načinu ubrzanja mreže. Gledajući MobileNet modele zaključeno je da bi se koncepti poput depthwise seperable konvolucije, inverzne rezidualne strukture, ekspanzijskih slojeva i „squeze and excitation“ mogli iskoristiti za ubrzanje prethodno opisanog Mask R-CNN modela. Također su proučeni benefiti jednostadijskih detektora implementiranih unutar RetinaNet-a i RetinaMaska te njihov utjecaj na performanse sustava. Na kraju je implementiran jednostavan sustav za segmentaciju objekata od pozadine, te su predložena dva sustava Efficient MaskRCNN i Efficient RetinaMask, koji bi potencijalno poboljšali performanse sustava i možda omogućili izvođenje na uređajima nižih performansi.

Abstract (english)

This paper focuses on finding ways to speed up the image segmentation process so that it can be used for video segmentation. First The R-CNN system (and its later iterations) for object detection and the Mask R-CNN for instance segmentation were investigated and by studying these systems and assessing their positive aspects and shortcomings, we turned to how to accelerate the network. Looking at MobileNet models, it was concluded that concepts such as depthwise seperable convolution, inverse residual structure, expansion layers and "squeze and excitation" could be used to accelerate the previously described Mask R-CNN model. The benefits of single-stage detectors implemented within RetinaNet and RetinaMask and their impact on system performance were also studied. Finally, a simple system for segmenting foreground form background objects was implemented and two systems, Efficient MaskRCNN and Efficient RetinaMask, were proposed which could potentially improve system performance and make the system viable on more low power devices.

• Keywords: R-CNN; Mask R-CNN; MobileNet; segmentacija; detekcija; RetinaNet; RetinaMask; konvolucija
• Keywords (english): R-CNN; Mask R-CNN; MobileNet; segmentacija; detekcija; RetinaNet; RetinaMask; konvolucija
• Language: croatian
• URN:NBN: urn:nbn:hr:168:045845
• 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: 2023-01-13 15:57:05