Regarding the serialisation of the model: in my opinion the size could be much compressed when using truncated random forest parameters without any expected loss of accuracy. However, I did not find a native way in opencv yet on how to do that.
In a possible next iteration, there are many points to work on, e.g.:
Swapping underlying comparators to the latest generation and hence improving the training data selection
Consider other interesting parameters in the GridSearch (as mentioned previously)
Optimising model complexity - performance
Improving training and model selection procedure
Improving by using other/more recent ML approaches
Using another framework with more flexibility, other than opencv for training
Regarding the serialisation of the model: in my opinion the size could be much compressed when using truncated random forest parameters without any expected loss of accuracy. However, I did not find a native way in opencv yet on how to do that.
In a possible next iteration, there are many points to work on, e.g.:
(Source: E-mail exchange with Daniel Hartung)