LabLiquidVision

The LabLiquidVolume dataset includes 5,451 images of liquids in laboratory containers. The images were taken using an Intel RealSense D415 camera in different environments in the automation laboratory and various research laboratories at the Novo Nordisk R&eD site in Måløv, Denmark. The ground truth of the liquid volume was measured using a Mettler Toledo XSR2002S balance with an accuracy of ± 0.5 mL. Twelve of the most common research laboratory containers, including the consumables used in cell culture processes, were selected. Various liquid volumes (3 - 600 mL), distances to the containers (50 - 600 mm), backgrounds, and camera angles were used. However, all images were taken from above so that the surface of the liquid is visible. In consultation with laboratory scientists, three different liquid colors were chosen: red, green, and blue. Besides transparent liquids, these are the most frequent colors of liquids in laboratory experiments. For the final dataset, the content is extended by the output of the segmentation and depth estimation model using the RGB images as input. This includes the segmented liquid and vessel depth maps, the liquid and vessel segmentation masks (each as .png and .npy files), and the unsegmented depth maps of liquid and vessel. The vessels in the images in the dataset have an average maximum volume of 178 mL and an average fill proportion of 48%. 31% percent of the samples contain green, 57% red, and 12% blue liquid.

Using this dataset, we propose a vision-based liquid volume estimation using a novel two-step Convolutional Neural Network (CNN) architecture. In the first step, a single RGB input image is processed by the first CNN to predict the segmentation and depth of the transparent container and the containing liquid. For the training of the first step, we benefit from existing datasets targeting transparent containers such as TransProteus and Vector-LabPics. These intermediate predictions are further processed by a second CNN to give an estimate of the liquid volume in the container. For the training of the second network, the proposed new dataset LabLiquidVolume is used.