BubDepth: A neural network approach to three-dimensional reconstruction of bubble geometry from single-view images

Abstract

The study of bubbly flows relies on the extraction of bubble information in experiments. Extraction with image processing based on images taken by high-speed cameras is a commonly adopted approach. Current methods mostly deal with silhouettes, abandoning the grayscale information in the images. In this paper, we propose BubDepth, a workflow that utilizes grayscale information and automatically reconstructs rough 3D shapes of one side of the bubbles from single-view images. The workflow consists of two parts: segmentation and depth inference. A neural network is used to recognize bubbles and masks in the segmentation part. The following depth inference network computes a relative depth map for each mask, describing the 3D shapes of one side of bubbles. The neural networks are trained using a dataset generated by computer graphics techniques. The image generator can create synthetic images of scenes labeled with 3D shape information of bubbles. BubDepth is a novel method for the 3D reconstruction of bubble shape based on single-view images. It achieved accurate results for synthetic images and could produce convincing predictions in the tests for real images.