Awesome article, Fabio! I liked the idea or comparing the relative merits of models by using the performance achieved as well as costs incurred. I would love to see this analysis extended to some of the well known neural network architectures like Faster-RCNN, YoLo, BERT, etc. It would also be easier to interpret the results with metrics like F-1 score instead of average precision for me personally. Keep up the good work :D