Colour videos with depth: acquisition, processing and evaluation

Abstract

<p>The human visual system lets us perceive the world around us in three dimensions by integrating evidence from depth cues into a coherent visual model of the world. The equivalent in computer vision and computer graphics are geometric models, which provide a wealth of information about represented objects, such as depth and surface normals. Videos do not contain this information, but only provide per-pixel colour information. In this dissertation, I hence investigate a combination of videos and geometric models: videos with per-pixel depth (also known as RGBZ videos). I consider the full life cycle of these videos: from their acquisition, via filtering and processing, to stereoscopic display.</p><p>I propose two approaches to capture videos with depth. The first is a spatiotemporal stereo matching approach based on the dual-cross-bilateral grid – a novel real-time technique derived by accelerating a reformulation of an existing stereo matching approach. This is the basis for an extension which incorporates temporal evidence in real time, resulting in increased temporal coherence of disparity maps – particularly in the presence of image noise.</p><p>The second acquisition approach is a sensor fusion system which combines data from a noisy, low-resolution time-of-flight camera and a high-resolution colour video camera into a coherent, noise-free video with depth. The system consists of a three-step pipeline that aligns the video streams, efficiently removes and fills invalid and noisy geometry, and finally uses a spatiotemporal filter to increase the spatial resolution of the depth data and strongly reduce depth measurement noise.</p><p>I show that these videos with depth empower a range of video processing effects that are not achievable using colour video alone. These effects critically rely on the geometric information, like a proposed video relighting technique which requires high-quality surface normals to produce plausible results. In addition, I demonstrate enhanced non-photorealistic rendering techniques and the ability to synthesise stereoscopic videos, which allows these effects to be applied stereoscopically.</p><p>These stereoscopic renderings inspired me to study stereoscopic viewing discomfort. The result of this is a surprisingly simple computational model that predicts the visual comfort of stereoscopic images. I validated this model using a perceptual study, which showed that it correlates strongly with human comfort ratings. This makes it ideal for automatic comfort assessment, without the need for costly and lengthy perceptual studies.</p><p>RGBZ videos, temporally coherent stereo matching, time-of-flight sensor fusion, stereoscopic viewing comfort</p>