Finite Aperture Stereo

Abstract Multi-view stereo remains a popular choice when recovering 3D geometry, despite performance varying dramatically according to the scene content. Moreover, typical pinhole camera assumptions fail in presence of shallow depth field inherent macro-scale scenes; limiting application larger scenes with diffuse reflectance. However, defocus blur can itself be considered useful reconstruction cue, particularly view-dependent materials. With this mind, we explore complimentary nature and cues context multi-view reconstruction; propose complete pipeline for modelling from finite aperature that encompasses image formation, calibration stages. As part our evaluation, an ablation study reveals how each cue contributes higher observed over range complex materials geometries. Though lesser concern large apertures, effects noise are also considered. By introducing pre-trained deep feature extraction into cost function, show step improvement per-pixel comparisons; as well verify cross-domain applicability networks using largely in-focus training data applied defocused images. Finally, compare number modern methods, demonstrate use both leads significant increase across several synthetic real datasets.