Building 3D models of objects from color images with next best view planning

Rekonstrukcija geometrije iz barvnih slik je eden izmed klasičnih problemov računalniškega vida. Kvaliteta pridobljenega 3D modela je močno odvisna od zajetih slik. Ročno zajemanje je lahko dolgotrajno opravilo, pri katerem želimo s slikami doseči ustrezno natančnost in pokritost modela. Uporabnik, ki med zajemanjem nima nobene povratne informacije o ustreznosti slik, s težavo upošteva vse predpostavke algoritmov, kar je lahko vzrok za neuspešno rekonstrukcijo, ponovno zajemanje pa je lahko zelo drago ali celo nemogoče. V tem delu smo se osredotočili na razvoj postopka in programske opreme, ki podpira celoten proces rekonstrukcije. Za vsako zajeto sliko dobi uporabnik sprotno informacijo o njeni ustreznosti ter oceni kvalitete trenutnega 3D modela. Razvili smo tudi novo metodo za načrtovanje naslednjih pogledov, ki sistematično izboljšajo kvaliteto rekonstrukcije. Metoda temelji na novi meri za oceno kvalitete 3D modela, za katero pokažemo, da je njen koeficient linearne povezanosti z dejansko natančnostjo boljši od obstoječe mere. Pokažemo tudi, da je rekonstrukcija z našo metodo načrtovanja naslednjega pogleda primerljiva in v nekaterih primerih boljša od rekonstrukcije z enakomerno postavljenimi kamerami po navidezni polkrogli.Reconstruction of geometry from RGB images is one of the classic computer vision problems. The quality of the produced 3D model heavily depends on the input images. Manual image acquisition can be a lengthy process with which we want to attain the desired accuracy and completeness of the model. A user that has no feedback about the suitability of the images during the acquisition process can have difficulties complying with assumptions of the algorithms, which can result in an unsuccessful reconstruction. Additional image acquisition can be expensive or even impossible. In this work, we focus on development of a system and software that support the entire reconstruction process. The user gets online information about the adequacy of every captured image and an estimate of quality for the current 3D model. We also present a novel method for next best view planning, that systematically improves the quality of reconstruction. The method is based on a new quality measure. We show that the linear correlation coefficient between our measure and accuracy is better than that of the existing measure. We also show that the reconstruction obtained by the next best view planning is comparable and in some cases better than reconstruction with evenly spaced camera configuration in the shape of a hemisphere