CHS: Small: Audio-Visual Reconstruction for Immersive Virtualized Reality

点击翻译按钮获取中文摘要

Maintaining the sense of presence is a major challenge in immersive virtual environments. An important aspect of immersion is the feeling of cohesiveness between different senses, including the visual and auditory; for example, an object that looks like wood should also sound like wood. Sound synthesis can improve a user's sensory cohesion when interacting with objects, but it requires accurate real-world material parameters. While much prior work in computer vision has focused on acquiring the geometric shape and visual characteristic of objects, the resulting point-clouds and images can assist in more accurate recovery of audio parameters for sound synthesis, along with acoustic scattering and absorption properties for sound rendering and propagation. A hypothesis of this research is that, conversely, auditory metrics can also assist in determining an object's geometry, including holes and occlusions, in a manner analogous to sonar detection (but of course 3D geometry reconstruction is far more challenging than object detection). The audio-visual reconstruction enabled by this project will have broad impact across many domains, including assistive technology for persons who are visually impaired, multimodal human-centric interfaces, immersive teleconferencing, rapid prototyping of acoustic spaces for urban planning, structural design, and noise control, to name just a few. Project outcomes including scientific advances and software systems will be disseminated through websites, publications, workshops, community outreach, and other professional events.This project explores a novel paradigm of audio-visual reconstruction of real-world scenes, where audio cues are used to guide the classification of objects and materials for 3D geometry reconstruction. At the same time, the visual information can be used to initialize and accelerate the identification of acoustic material parameters. Some of the major research challenges that will be addressed include audio-guided 3D model reconstruction, design of audio-visual neural networks for material and object identification, learning-based acoustic material classification of a large physical or virtual space, and optimization-based acoustic material refinement using geometric and wave-based methods. Perceptually-grounded evaluation and validation of the new methods and applications will be performed.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.