The main issue we faced was speed and performance of the face-tracking, especially on older iPads. Since visual computations tend to be expensive, the lack of processing power on old tablet models would result in sluggishness of the interface and overheating, especially when processing multiple faces simultaneously. We moved heavy processes to the cloud by using Azure Face API for emotion and face detection, as well as face recognition, and performed object tracking offline periodically using the iOS Vision API.

Another issue would be the recognition accuracy for young children. We filtered, normalised and tagged a good amount of photos of participating pre-schoolers to train the model on Azure Face. We also proposed that newly taken and normalised profiles should be used in training and updating the model to progressively keep up with the children's growth.

A key concern would be to track faces of active children who tend to move about in the frame. We improved our tracking fidelity by toggling between authoritative face detection using Azure Face API and auxillary object tracking on Vision API. With object tracking being less computation-intensive, we could afford higher refresh rates and more reliable tracking with relatively lesser performance trade-off.