Neural Numerical Estimation of Liquid Volume in Transparent Containers via a Single RGB-D Image

Measuring liquid volume can be perilous and in-efficient under some circumstances. To address this problem, we present a new approach to measuring liquid volume in transparent containers from a single RGB-D image. We propose a novel 10-bit binary encoding of the volumetric value with a corresponding numerical binary-code cross-entropy loss function, i.e., NBCCE loss. We achieve better volumetric prediction than previous methods. Then, We construct a fully convolutional anchor-free volumetric prediction network and embed NBCCE loss into this network. To verify our method, we build an RGB-D image dataset of transparent containers filled with varying amounts of liquid. In experiments, we incorporate our binary encoding and NBCCE loss into a group of classical network structures. Our method results in significantly lower estimation error than L1, smooth L1 or MSE with direct floating number encoding, and the fine-grained classification [1], [2]. Furthermore, our method also outperforms the existing fluid volume estimation method using the equivalent dataset.