Semantičko prognoziranje regresijom pomaka značajki

master's thesis

• Title (english): Semantic forecasting by regressing feature motion
• Author: Iva Sović
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Zoran Kalafatić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2021-07-16, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Ovaj rad se bavi kratkoročnim i srednjoročnim semantičkim prognoziranjem budućnosti koje predviđa semantičku mapu određenog budućeg trenutka. Definiraju se i opisuju F2F i F2M pristupi prognoziranju. Implementirana su dva glavna modela kao okosnica za testiranje izložene teorije: \emph{single-frame} model za semantičku segmentaciju te F2MF model za prognoziranje. Predstavljen je novi način učenja \emph{single-frame} modela tako da mu je omogućeno mješovito učenje s ili bez preskočnih veza. F2MF model je nadograđen tako da može prognozirati preskočne veze i iskoristiti njihove vrijednosti za poboljšanje točnosti prognoziranja. Za kratkoročno predviđanje postignuta je točnost od 69.6 postotnih bodova, a za srednjoročno predviđanje točnost od 57.4 postotnih bodova. Oba dobivena rezultata su usporediva s rezultatima u literaturi, točnije bolji su za 0.9 odnosno za 0.6 postotnih bodova u usporedbi s odgovarajućim eksperimentima.

Abstract (english)

This work addresses the topic of short-term and mid-term semantic forecasting. The definitions of the F2F and the F2M forecasting approaches are presented. In the scope of this work, two main models were implemented and used as backbones for testing the presented theory: a single-frame model for semantic segmentation and a mixed model for forecasting. We presented a novel learning approach for the single-frame model where learning can be conducted using all, some or no skip connections. The F2MF model was extended so that it can both forecast the skip connections and also use them to improve the accuracy of its forecasting. The tested model achieved mIoU accuracies of 69.6 percentage points and 57.4 percentage points for short-term and mid-term forecasting, respectively. This constitutes an accuracy improvement compared to the current state-of-the art by 0.9 and 0.6 percentage points, respectively.

• Keywords: semantičko prognoziranje; single-scale model; F2M; F2F; F2MF; mješovito učenje; prostorno izobličavanje; optički tok; kratkoročno predviđanje; srednjoročno prognoziranje
• Keywords (english): semantic forecasting; single-scale model; F2M; F2F; F2MF; mixed learning; warp; optical flow; short-term forecasting; mid-term forecasting
• Language: croatian
• URN:NBN: urn:nbn:hr:168:116971
• 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: 2023-01-15 11:14:43