Firstly, we have identified the size of the retrain models. As shown in the following screenshot, Mobilenet V1 requires only 17.1 MB (for both use cases), which is than one-fifth of Incentive V3 (92.3 MB), and this model can be easily deployed in resource-constrained IoT devices, including Raspberry Pi or smartphones. Secondly, we have evaluated the performance of the models. Two levels of performance evaluation have been done for the use cases: (i) dataset-wide evaluation or testing has been done during the retraining phase on the desktop PC platform/server, and (ii) an individual image or sample (real-life image) was tested or evaluated in the Raspberry Pi 3 environment:
