Capturing real-world dynamic objects using temporally-coded photography

Extracting high-quality 2D and 3D information from real-world objects plays an important role in surveillance, simulation, scientific analysis, virtual reality, entrainment, and other commercial and scientific applications. Such a capturing task should work not only for static and diffuse objects, but also for fast-moving and deforming objects as well as optically-complex objects. Today this task often requires expensive equipments or complicated hardware setup and calibration. This dissertation research aims to develop robust methods that capture high-quality models for real-world dynamic objects efficiently, accurately, and with low cost. One-frame methods are suitable for moving scenes but typically result in low-resolution models. The addition of temporal processing (e.g., multiple frames) achieves higher resolution and quality but is challenging for dynamic objects. In this dissertation, fundamental problems, which appear when using temporal processing to capture moving and deforming objects, are addressed. The key observation is that for certain types of object motion, multiple frames captured at different time instances can be registered together to a given time instance; enabling the use of more than one image, each with different imaging parameters, to capture the virtually static state at the given instance. According to this observation, this research develops a series of methods for acquiring high-resolution time-varying 3D geometric models for moving and deforming objects using simple hardware infrastructures. In addition this dissertation also studies and proposes approaches that capture photographs that are free of motion blur for fast-moving objects. Finally, technique that reconstructs the 3D geometry for optically-complex scenarios is also presented. Collectively, the approaches developed in this dissertation lead to the ultimate goal of high-quality capture of moving and deforming objects