Context Counting objects in images is a task with many applications. The idea was to create a simple data set to test some ideas regarding this task. The data set has a large number of images compared to others for this similar task, but also allows for quick training in a GPU enabled Kernel. Your task is to correctly count the number of objects in each image, divided in three types: Blue circles, green squares and red squares. They are semi-transparent and allowed to overlap, and each has some specific properties about their size, frequency distribution, among others. The images are presented both in their original form and as numpy arrays (provided as pkl objects) that can be loaded and readily used on Keras. Content The data set contains 50000 images (42500 training, 7500 test). Within these, there are slightly over 300000 shapes. The labels are lists of 3 values, namely the quantities of blue circles, green squares and red squares. Different architectures and evaluation metrics can be tested. Some baseline results have been uploaded as a Kernel. Inspiration Blue circles are the most frequent shape. Green squares have the most variability in size. Red squares are the least frequent shape, but are constant in size. Can we guess which one of these will yield the most miscounts for a given counting algorithm? Since the data set is relatively simple, can we get the number of miscounts down to 0? How would these questions change if we added more colors, more types of shapes, a higher number of shapes per image, among other challenges? The hope is that this data set, or future incremented versions of it, can help testing some of these concepts and others regarding counting algorithms. Suggestions on how to improve the data set are greatly appreciated.