Nenadzirano učenje modela za monokularnu procjenu dubine

master's thesis

• Title (english): Unsupervised learning of models for monocular depth estimation
• Author: Ivan Bilić
• Mentor: Siniša Šegvić (mentor)
• Committee member: Siniša Šegvić (predsjednik povjerenstva)
• Committee member: Marko Čupić (član povjerenstva)
• Committee member: Ivan Petrović (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

Monokularna procjena dubine loše je postavljen problem kojeg karakterizira nedostatak geometrijskih ograničenja, a primjena dubokog učenja pokazuje se plodonosnom u računalnom vidu. Pristup zadatku podrazumijeva nenadzirano učenje dubokih konvolucijskih neuronskih mreža za procjenu dubine i relativne poze kamere, temeljenih na široko rasprostranjenoj arhitekturi U-Net. Pojedine induktivne pristranosti te korisna geometrijska svojstva - poput rijetke povezanosti i ekvivarijantnosti s obzirom na pomak, razlog su zbog kojeg spomenute arhitekture predstavljaju prikladan alat za rješavanje navedenog zadatka. Matematički okvir za nenadzirano učenje monokularne procjene dubine proizlazi iz modela kamere i geometrije nastanka slike, a svodi se na zadatak rekonstrukcije. Kod monokularnog pristupa, posebno se ističe problem skale te potreba za modelom za procjenu relativne poze kamere - dva povezana problema kojih nema kod stereo pristupa. Kvalitativni rezultati pokazuju pojedine učinke učenja i ugađanja modela na skupu KITTI te na internom skupu BIH. Arhitekturalne značajke poput procjene aleatorne nesigurnosti, rekonstrukcije značajki i mehanizma pozornosti, pokazuju se kao kompetitivne. Eksperimenti također pokazuju da se performanse monodepth2 modela na KITTI Eigen testnom skupu mogu unaprijediti pažljivijim odabirom vrijednosti pojedinih hiperparametara te upotrebom transponirane konvolucije za naduzorkovanje. Svi eksperimenti su temeljeni na monodepth2 modelu.

Abstract (english)

Monocular depth estimation is an ill-posed problem that suffers from lack of geometric constraints, whilst deep learning has already proven itself for a variety of computer vision tasks. The approach consists of training deep convolutional neural networks deployed as popular and widely used U-Net architecture. These tools are known for possessing desirable inductive biases (e.g. sparse connectivity) and geometric properties (e.g. translational equivariance). Unsupervised learning for monocular depth estimation is enabled by a mathematical framework that arises from single view geometry and the pinhole camera model; the neural networks are trained to perform warping-based view synthesis. The "scale problem" and the need for relative camera pose estimation model are the key differences between monocular and stereo approach. Qualitative results show the effects of training and fine-tuning on KITTI and internal BIH datasets. Architectural features such as aleatoric uncertainty estimation, multi-scale feature reconstruction and attention mechanism are also competitive. Experiments also show that monodepth2 performance on KITTI Eigen test set can be improved by choosing different hyperparameter values and using transposed convolution for upsampling. All experiments are based on monodepth2 model.

• Keywords: monokularna procjena dubine; nenadzirano učenje; samonadzirano učenje; konvolucijske neuronske mreže; duboko učenje; model rupičaste kamere
• Keywords (english): monocular depth estimation; unsupervised learning; self-supervised learning; convolutional neural networks; deep learning; pinhole camera model
• Language: croatian
• URN:NBN: urn:nbn:hr:168:957646
• 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 11:49:23