The Virtual Window Wall
Author
Pedersen, Casper
Term
4. term
Publication year
2013
Abstract
Denne masteropgave præsenterer design og implementering af prototypen Virtual Window Wall, et billedbaseret renderingssystem, der beregner en virtuel visning af en scene fra perspektiver mellem flere referencekameraer og kan styres af brugerens hovedposition. Prototypen er udviklet i C++ og udnytter OpenCL til at køre centrale algoritmer parallelt på GPU’en. Arbejdsgangen omfatter optagelse af kalibreringsbilleder, kamerakalibrering og generering af rektificeringskort, efterfulgt af rektificering af et kamerapar og estimering af disparitets-/dybdemaps med flere implementerede metoder, herunder nye varianter. De virtuelle visninger dannes ved en ny metode, baglæns 3D-warping med disparitiesøgning, som udnytter den epipolare begrænsning til at bestemme lokal dybde for det virtuelle billede. Test viste perceptuelt realistiske resultater med mindre artefakter inden for et fornuftigt område omkring referencekameraerne; den mest markante fejlskyldtes okklusioner. GPU-acceleration muliggjorde interaktiv manøvrering af det virtuelle kamera med omkring 17 billeder pr. sekund, men den beregningstunge cost space-aggregationsfase var en flaskehals, selv om den stadig kørte cirka 17 gange hurtigere på GPU end på CPU.
This master’s thesis presents the design and implementation of the Virtual Window Wall prototype, an image-based rendering system that synthesizes virtual views of a scene between multiple reference cameras and can be controlled by the user’s head position. The prototype is implemented in C++ and uses OpenCL to execute key algorithms in parallel on the GPU. The pipeline captures calibration images, performs camera calibration and generates rectification maps, then rectifies a camera pair and estimates disparity/depth maps using several implemented methods, including novel variants. Virtual views are produced with a new method, backward 3D warping with disparity search, which leverages the epipolar constraint to infer local depth for the virtual image. Tests showed perceptually realistic results with minor artifacts within a reasonable region around the reference cameras; the most noticeable errors were due to occlusions. GPU acceleration enabled interactive virtual camera control at about 17 frames per second, though the computationally heavy cost space aggregation step remained a bottleneck, even while running roughly 17 times faster on the GPU than on the CPU.
[This summary has been generated with the help of AI directly from the project (PDF)]
Keywords
Documents