Gigapixel imaging aims to capture large-scale dynamic scenes with both high resolution and wide FoV, which is indispensable to many research areas. However, gigapixel imaging associates with a large space-bandwidth product, and it is almost impossible to be achieved with conventional optical imaging systems restricted by the optical diffraction, geometric aberration, scattering medium, etc. In this thesis, we focus on designing adaptable, robust, and versatile parallel sensor arrays, leading to practical usable gigapixel imaging systems.

We first propose UnstructuredCam, consisting of one wide-FoV micro-camera for the panoramic view capturing and several unstructured long-focus micro-cameras for the local details capturing. This design liberates the overlapping requirements betw...[ Read more ]

Gigapixel imaging aims to capture large-scale dynamic scenes with both high resolution and wide FoV, which is indispensable to many research areas. However, gigapixel imaging associates with a large space-bandwidth product, and it is almost impossible to be achieved with conventional optical imaging systems restricted by the optical diffraction, geometric aberration, scattering medium, etc. In this thesis, we focus on designing adaptable, robust, and versatile parallel sensor arrays, leading to practical usable gigapixel imaging systems.

We first propose UnstructuredCam, consisting of one wide-FoV micro-camera for the panoramic view capturing and several unstructured long-focus micro-cameras for the local details capturing. This design liberates the overlapping requirements between adjacent local micro-cameras, enables content-adaptive unstructured sampling, and significantly improves the sensor utilization. After that, we build PANDA using our UnstructuredCam: the worlds first gigaPixel-level humAN-centric viDeo dAtaset focusing on large-scale, human-centric, long-term, and multi-object visual analysis. Representative computer vision tasks are benchmarked and a novel task, interaction-aware group detection, is studied.

Then, we present multiscale-VR: multiscale gigapixel 3D panoramic videography for virtual reality. The system comprises scalable cylindrical-distributed global stereo micro-cameras and unstructured local micro-cameras. The former ones are used to cover a 360-degree FOV and estimate the panoramic depth map. The latter ones are adapted for local high-resolution video streaming and depth refinement, where a novel cross-resolution trinocular depth estimation algorithm is presented to merge geometric and semantic information.

Due to the high-resolution demand of biology studies, we also apply our design to Giga-PACT: a photoacoustic computed tomography system towards volumetric gigapixel imaging using massively parallel ultrasonic sensor arrays. Our system is comparable to the highest clinical standard (7 T MRI scanner) under a specific depth range on subjects with craniectomies in functional human brain imaging studies, serves as a good alternative for MRI incompatible subjects at a much lower cost with potential portability.