END-TO-END LITTER DETECTION USING A TINY HYBRID DATASET

Abstract

Object detection typically requires vast amounts of data to be trained effectively. As the performance of object detection networks begin to plateau, more attention is being paid to the quality of data over the quantity of data. This thesis demonstrates how a tiny hybrid dataset of 10 to 20 images can be synthesised to train a model that effectively detects common litter around the built environment using a smartphone. The results and observations from this thesis show that a model trained from a tiny dataset of real-world images may seem to perform the best during training, but is outperformed by its hybrid counterpart in reality when applied back to the field.