Sometimes it's unavoidable or even desirable to include the human in tuning the output from the model. In these instances, the model is used to assist the user rather than completely automating the task. A few approaches that could be employed for the project in this chapter include the following: 

• Provide an intermediate step where the user can tidy up the masks. By this, I mean allowing the user to erase parts of the mask that have been incorrectly classified or are unwanted by the user.
• Present the user with a series of frames and have them select the frames for composition.
• Present the user with variations of the final composited image and have them select the one with the most appeal.

Another related concept is introducing human-in-the-loop machine learning. This has a human intervening when the model is not confident in its prediction, and it passes the responsibility over to the user for classification and/or correction. The amendments from the user are then stored and used for training the model to improve performance. In this example, we could let the user (or crowd-source this task) segment the image and use this data when re-training the model. Eventually, given sufficient data, the model will improve its performance relevant to the context it is being used in.

I hope this section highlighted the importance of handling uncertainty when working with machine models and provided enough of a springboard so that you can approach designing intelligent applications from the perspectives outlined here. Let's now conclude this chapter by reviewing what we have covered.