Estimating the level of water in drinking glasses is important for monitoring the water in-take of a person. We provide a solution to this task by training deep neural networks on images of drinking glasses with different levels of water in them. We generated new data sets comprising 5500 real-world and 190350 synthetic images of drinking glasses filled with different levels of water. The real-world images were captured using various cameras and in different environments. The synthetic images were rendered using the 3D modeling software Blender. We found that the number of real-world images in the training set directly impacted the outcome and the best accuracy of 0.76 for 40 and 0.85 for 8 classes was achieved when all 5500 real-world images were used for training. In case of reduced number of only 2080 real-world images in the training set, complement of synthetic images greatly improved the best accuracy from 0.44 to 0.77 for 40 classes. Out of the five deep neural networks trained, Vgg16, Inception_v4, and Resnet_v2_200 performed best closely matching each other’s classification accuracies; whereas, Mobilenet_v2 and PNASnet_mobile performed poorly.