Efficient Bounding Box Selection for Object Localization

Autonomous edge systems require a high level of accuracy with a small energy footprint for integration into real time applications. This paper proposes a novel neural network, the Bounding Box Vision Transformer (BB-ViT), designed to provide efficient localization of a single object of interest. BB-ViT is a vision transformer built for both speed and energy efficiency at only 40,092 parameters, it is capable of achieving mean intersection of union scores above 70% for five diverse classes of the Caltech101 dataset. On the Airplane class the network received a score of 92.52% mean IoU and MSE of 0.00042, equating to superior confidence of target localization.