This study applies deep neural networks to the task of detecting semi-transparent drinking glasses on images captured in specific as well as generalised surroundings. Detecting semi-transparent objects is a major challenge in the fields of computer vision and machine learning. In this study, we use a mix of 40000 synthetically-generated photo-realistic images and 8000 real-world images of multiple types of drinking glasses placed in various ambient surroundings. Two separate test sets were generated to compare the performance of five deep neural networks when trained using different proportions of synthetic and real-world images. The test sets include 800 images taken in typical household environments and 1000 images taken in an aged care facility. The training was performed on five deep neural networks using Tensorflow’s object detection API. The results show that models initially trained on large quantities of synthetic data, and further trained using a small number of real-world images produce better detection accuracy than models trained on large quantities of real-world images only. Furthermore, the detection accuracy is higher if we use a model pre-trained on synthetic images and provide additional training using a very small number of real-world images taken in the same surroundings as the test set. Among the five models used in our study, Faster R-CNN inception, Faster R-CNN resnet101, and R-FCN resnet101 performed best, closely matching each other’s detection accuracies. On the other hand, SSD inception and SSD mobilenet performed considerably worse compared to other networks on our data.