Humans can use ambient sounds like ventilation noise to ticking clocks to understand 3D scene structure. A recent paper on arXiv.org investigates whether these sounds can be used for multimodal self-supervised learning.
Sound waves. Image credit: mtmmonline via Pixabay, CC0 Public Domain
The researchers collected a dataset of “in-the-wild” audio recordings from quiet, indoor scenes typical of what a robot would encounter when solving navigation tasks. Each sound is paired with a corresponding recording from an RGB-D sensor, which provides a visual signal and pseudo ground-truth depth. An experimental study of depth estimation was conducted using the dataset. It is demonstrated that audio can be used to estimate the distance to nearby walls.
The suggested model can be used as part of a simple robotic navigation system, in which a wheeled robot moves along a wall using ambient audio cues. It is also shown that audio-visual recordings can provide useful self-supervision for depth estimation tasks.
From whirling ceiling fans to ticking clocks, the sounds that we hear subtly vary as we move through a scene. We ask whether these ambient sounds convey information about 3D scene structure and, if so, whether they provide a useful learning signal for multimodal models. To study this, we collect a dataset of paired audio and RGB-D recordings from a variety of quiet indoor scenes. We then train models that estimate the distance to nearby walls, given only audio as input. We also use these recordings to learn multimodal representations through self-supervision, by training a network to associate images with their corresponding sounds. These results suggest that ambient sound conveys a surprising amount of information about scene structure, and that it is a useful signal for learning multimodal features.
Research paper: Chen, Z., Hu, X., and Owens, A., “Structure from Silence: Learning Scene Structure from Ambient Sound”, 2021. Link: https://arxiv.org/abs/2111.05846