Fast R-CNN Implementation Paper: https://arxiv.org/pdf/1504.08083.pdf
The Fast R-CNN research paper recommends using 32 positive ROIs and 96 negative ROIs for each image. However, the selective search algorithm that was used for this implementation produced far less positive ROIs. In most cases, it produced 3 or 4 positive RoIs per image. Therefore, each image had substantially less ROIs to learn from as the number of negative ROIs used was always 3x the amount of positive ROIs produced. This meant we needed far more epochs than the 40000 that was recommended. This model was only trained for 30000 epochs.
Run time is approximately 2.5 to 3 seconds.
