Izlučivanje kandidata u dvoprolaznim lokalizacijskim modelima

undergraduate thesis

• Title (english): Candidate extraction in two-stage localization models
• Author: Leo Pleše
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Mladen Karan (član povjerenstva)
• Committee member: Josip Šarić (član povjerenstva)
• Granter: University of Zagreb; Faculty of Electrical Engineering and Computing; Zagreb
• Defense date and country: 2020-07-09, Croatia
• Scientific / art field, discipline and subdiscipline: TECHNICAL SCIENCES; Computing

Abstract

Postoje mnogi modeli lokalizacije objekata: mi smo odabrali RPN. RPN je dvoprolazni lokalizacijski model duboke konvolucijske neronske mreže koji paralelno predviđa klasu objekta u razred objekt/pozadina te položaj i dimenzije okvira objekata. Pritom se oslanja na model FPN koji radi i prolaz konvolucijskom mrežom u suprotnom smjeru čime izlučuje još značajki spojivši lateralnim vezama mape značajki tih prolaza – lokalne informacije iz prvog i semantički vrednija svojstva iz drugog. Za osnovu FPN-a odabrali smo arhitekturu ResNet-34 koja rješava problem degradacije točnosti na skupu za učenje uslijed prevelike dubine konvolucijske mreže. Proučili smo ove modele i implementirali RPN. Naučili smo model na više slika iz podatkovnog skupa PASCAL VOC 2012 pojedinačno i eksperimentalno metrikom AR1000 potvrdili da učinkovito označava interesne regije okvirima kao rezultat učenja pozicija i dimenzija okvira radi još bolje prilagodbe objektu. Na kraju smo algoritmom NMS uspješno odabrali okvire koji se prema IoU najbolje preklapaju s točnima.

Abstract (english)

There are many object localization objects: we have chosen RPN. RPN is a two-stage convolutional neural network localization model which in parallel predicts an object's class into object/background as well as position and dimension of its bounding boxes. Thereat it relies on FPN model, which does an additional pass through convolutional network in the opposite direction, which extracts more features by having lateral connections between the feature maps from the two passes – local information from the first and semantically more valuable features from the second one. As for the FPN groundwork we have chosen ResNet-34 architecture, which solves the training accuracy degradation problem appearing as a result of too large convolutional network depth. We have studied these models and implemented RPN. We have trained the model on multiple images from the PASCAL VOC 2012 dataset individually and by using AR1000 for the baseline experimentally confirmed it efficiently labels regions of interest with bounding boxes as a result of learning anchor positions and dimensions with aim of even better adjustment to the object. At the end, using NMS algorithm we have successfully selected the anchors which according to IoU have the largest overlap with the ground truth ones.

• Keywords: RPN; Mask-RCNN; FPN; ResNet-34; konvolucijske neuronske mreže; duboko učenje; diferencijabilno programiranje; strojno učenje; umjetna inteligencija; podatkovni skup PASCAL VOC 2012; AR1000; lokalizacija objekata; detekcija objekata; segmentacija objekata; dvoprolazni lokalizacijski modeli; IoU; interesne regije; okviri; točni okviri; NMS
• Keywords (english): RPN; Mask-RCNN; FPN; ResNet-34; convolutional neural networks; deep learning; differentiable programming; machine learning; artificial intelligence; PASCAL VOC 2012 dataset; AR1000; object localization; object detection; object segmentation; two-stage localization models; IoU; regions of interest; anchors; bounding boxes; NMS
• Language: croatian
• URN:NBN: urn:nbn:hr:168:408498
• Study programme: Title: Computing; Study programme type: university; Study level: undergraduate; Academic / professional title: sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva (sveučilišni/a prvostupnik/ prvostupnica (baccalaureus/baccalaurea) inženjer/inženjerka računarstva)
• Type of resource: Text
• File origin: Born digital
• Access conditions: Closed access
• Created on: 2023-01-13 15:54:00