Implement SSD-MobileNet ArmNN workflow

[psyhtest]

Single-Shot Detector (SSD) is a popular approach for object detection. It can be paired with different "heads" e.g. MobileNet or ResNet used in MLPerf Inference.

While we have a TensorFlow (Python) implementation of SSD-MobileNet, the preference is to start with a TensorFlow Lite one.

[psyhtest]

First we need to create a reference TensorFlow Lite workflow. While we don't need to start with a quantized version, we may find the experience of going from tf_ssd_mobilenet_v1_quant to tflite_ssd_mobilenet_v1_quant useful. I am told this was achieved via the export_tflite_ssd_graph.py script, as detailed in this blog post.

As I understand, one of the challenges is that TFLite 0.1.7 does not support control flow. This is also the case for ArmNN, so a similar workaround should work.

[psyhtest]

Work-in-progress.