Google provides a number of models that are included in the object's detection API. The models include convoluted neural network models and simple models designed to operate on devices with standard specifications.
Earlier this week Google also announced its line of introduction to the "lightweight" version of MobileNets. This model can handle a number of tasks such as object recognition, face recognition, and landmark recognition.
Admittedly, the smartphone is currently not equipped with computing resources equivalent desktop and server-based settings. Impact, developers have two options.
Both options are machine learning models that can run on cloud computing platforms, but have latency limitations and require an internet connection. While the other alternative is to simplify the model itself.