Deep Learning for Free-Viewpoint Video in Sports and Immersive VR Experiences

点击翻译按钮获取中文摘要

The increase in popularity of virtual reality (VR) and augmented reality (AR) experiences has given rise to a need for high-quality immersive content. Until now these experiences have predominantly featured artist-made content, but the process of creating realistic models and textures this way is challenging and time-consuming. There has been recent uptake in the use of vision-based methods for creating virtual content from real images, including Free-Viewpoint Video (FVV). FVV allows us to reconstruct real-world scenes and present them in a virtual medium. Most FVV methods require large volumes of geometry and texture data, and can require significant processing capabilities for real-time rendering. On top of these challenges, sports scenes are an especially difficult subject for FVV due to the unconstrained environment, sizeable performance volume andlarge number of people. The overall focus of this work is to investigate methods for the offline processing and real-time rendering of FVV of highly dynamic scenes such as sports. FVV pipelines employ a geometric representation of the scene data, which is textured using camera images. The geometry is usually obtained using 3D reconstruction from multiple camera viewpoints - either general or model-based. General multi-view reconstruction methods are not reliant on prior information about the scene structure, and many have been applied to the generation of FVVcontent. Model-based reconstruction methods involve fitting a shape-model to real-world data to obtain a geometrical proxy of the scene. We believe that model-based methods can be employed in the production of FVV of sports and dynamic scenes, providing multiple benefits over existing methods:- Rather than a unique geometry per person and frame, a model-based representation requires only a set of model parameters plus the overhead of the body model. This means that modelbased FVV would be easily scalable to multiple people.- Since the output of model-based reconstruction is intrinsically consistent over time, video compression on the texture maps is more effective, further increasing compactness.- Model-based reconstruction is more robust to the errors that general reconstruction methods are prone to. Prior knowledge of human shape and pose embedded in the model assists in overcoming visual ambiguities present in the input data, providing a whole-body reconstruction without missing regions.- The model may allow finer detail to be inferred where other capture methods are insufficient, such as around the face and hands.- If the body model features an articulated skeleton, the reconstruction can be easily refined or reanimated by an artist.The merits of model-based reconstruction make it suitable for FVV, especially for sports scenes. This work focuses on employing it in the production of immersive content for VR and AR. The aim is to achieve real-time FVV rendering with temporally coherent representations for compact streaming of dynamic scene data. A particular focus will be on the application to sports, both for player performance and analytics, and the creation of mmersive content for VR and AR experiences. FVV production for sports scenes is an especially challenging application since the environment is uncontrolled. The associated difficulties include inaccurate calibration, uncontrolled illumination, poor segmentation, and very wide-baseline cameras.