Classifying Negative Objects With Neural Networks

The term object is generally only associated with something that can be sensed. However, the empty space between and around objects can also be considered as objects, which we will define as negative objects. While incredibly important in some instances, they have been neglected by the research community in comparison to their positive counterparts. Without a properly developed lexicon for them, negative objects are hard to discuss and describe. This thesis develops the starting point for a lexicon for negative objects, builds a publicly accessible dataset, and demonstrates that they can be identified within an image with the application of a machine learning algorithm. The neural network performs at an average precision of 74% when identifying a certain type of negative object (holes). This algorithm also shows promise in being able to differentiate between holes and tunnels.