Training consists of 3 steps (given AP values for full validation dataset):
- Training from MobileNet weights.
- Training from weights, obtained from previous step.
- Training from weights, obtained from previous step and increased number of refinement stages to 3 in network.