The first is improving the model. Of course, there may be limitations depending on the source of the model and dataset, but it's important to be able to understand ways in which the model can be improved as its output directly correlates to the quality of the user experience.

In the context of this project, we can augment the model using an existing pre-trained image classifier as the encoder, as mentioned earlier. This not only fast-tracks training, providing more opportunities to iterate, but also is likely to improve performance by having the model transfer existing knowledge from a more comprehensive dataset. 

Another is tuning the dataset that the model was trained on. A simple, and relevant, example of how the model can be improved can be seen by any image in which the user is holding an object (which has been labeled). An example of this can be seen in the following figure, in which the guitar is cropped from the person:

 

How you deal with this is dependent on the desired characteristics of your model. In the context of the application presented in this chapter, it would make sense to either perform multi-class classification, including objects normally held by people, or including them in the mask. 

Another common technique is data augmentation. This is where you artificially adjust the image (input) to increase variance in your dataset, or even adjust it to make it more aligned with the data for your particular use case. Some example augmentations include blurring (useful when dealing with fast moving objects), rotation, adding random noise, color adjustment - essentially any image manipulation effect that introduces nuances that you are likely to get in the real world.

Of course, there are many more techniques and tools to improve the model and data; here, our intention is just to highlight the main areas rather than delve into the details.