Duboki modeli za detekciju ključnih točaka osoba

master's thesis

• Title (english): Deep models for human key-point detection
• Author: Pavo Matanović
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Jan Šnajder (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2022-09-19, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Detekcija ključnih točaka osoba važno je područje računalnog vida sa zanimljivim
primjenama. Opisanim postupkom moguće je modelirati gotovo sve stupnjeve slobode
čovjeka. Prvi dio rada opisuje model Cascaded Pyramid Network koji je dvofazni model, u prvoj fazi se radi detekcija osoba na slikama dok se u drugoj na osnovi detekcije
osobe označuju ključne točke. Drugi dio rada opisuje preinaku modela SwiftNet tako
da detektira ključne točke. Ideja je bila napraviti model koji ima male memorijske zahtjeve da ga možemo trenirati na dostupnom GPU, a da ima što bolje performanse. Treniranje i evaluaciju problema traženja ključnih točaka osoba provodili smo na skupu
MS COCO 2017. Na ispitnom skupu model CPN postiže točnost od 65.4, odnosno
67.1 ako testiramo samo fazu traženja ključnih točaka. Preinačeni model SwiftNet
postiže točnost od ?.

Abstract (english)

Human key-point detection is important field of computer vision with interesting applications. Using the described method it is possible to model almost all degrees of freedom on human. The first part of paper describes two-phased model named Cascaded Pyramid Network, where the first phase is human detection while the second phase is key-point detection for detected human crops. The second part of paper describes modified SwiftNet model for key-point detection. Main idea was to make a model with small memory footprint that can be trained on available GPU, while trying to achieve good performance. Training and evaluation was conducted on MS COCO 2017 dataset. On test dataset CPN achieves accuracy of 65.4 and 67.1, respectively with human detector and with original human crops. Modified SwiftNet achieves accuracy of ?.

• Keywords: Detekcija ključnih točaka osoba; CPN; SwiftNet; MS COCO
• Keywords (english): Human key-point detection; CPN; SwiftNet; MS COCO
• Language: croatian
• URN:NBN: urn:nbn:hr:168:254160
• 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 09:29:39