A tile-based school classifier is a binary image classification based on Xception²⁰ pre-trained model from ImageNet²¹. The direct school detection model is an Object Detection model that we adapted from TensorFlow Object Detection API²² , specifically, we adapted the pre-trained SSD MobileNet ResNet101²³ Coco model. Both models were deployed on top of Kubeflow²⁴ and Kubernetes²⁵ that allow machine learning and cloud engineers to run model training and experimentations quickly and efficiently with TFJobs (Figure 7 and Figure 8). The Kubeflow is a tool that makes ML workflows on Kubernetes be deployed easier, simpler, portable, and scalable. Each model training and model experiment was recorded with TFJob YAML files, so it’s traceable. The best performing model is identified and selected with a set of the model evaluation metrics for tile-based school classifiers or direct school detection. For the tile-based school classifier, we used F1, precision, and recall scores as well as a ROC curve from model evaluation over test dataset. To evaluate metrics for the direct school detection model, we compute confusion matrices, F1 scores, mean average precision, and recall scores.

__________________________________