I implemented the neural baseline version). I anneal the weight of the reinforce term from 0 to 1 and presence prior probability from 0.5 to 0.01 and these stabilize training. Some localization examples:
Counting accuracy reaches around 92.0%. This is not as good as reported in the paper (over 98%), and this may be an issue of hyperparameters tuning.
