We used data collected from the CPPE-5 dataset by Dagli and Shaikh [1] for our purpose. Refer to Figure 1 for examples of images from this dataset. We found this dataset suitable because it consists of only 1029 labeled images, which we consider to be very small compared to traditional datasets like COCO [2] which consists of 328,000 images. We also find this dataset to be class-imbalanced, which makes it similar to what would be collected in practical scenarios. The dataset also consists of a large number of non-iconic images making sure all the images are real-life images unlike other datasets in this area.