Custom Object Detection with Deep Learning - Using TensorFlow Lite on Small Computers

Been learning convolutional neural network principles and deep learning for a few months. Here is a demonstration of training an existing deep learning network - SSD Mobilenet V2*, for custom object detection using TensorFlow framework (Using V2.6 at present). By using the method of 'transfer learning', I trained the network to detect tennis ball and geared DC motor, using a small datasets of 96 photos.
*Single-shot detect, a popular one-stage object detection deep neural network pipeline, which has a feature extraction front-end, and a backup that performs classification and localization of objects. For this instance, the feature extraction front-end is using Mobilenet V2 backbone ('backbone' means a sequence of neural-network layers).

In the video I show the comparison of running the model on TensorFlow Lite interpreter on a standard laptop with Intel Core i5 processor (no GPU) with 12GB of RAM and Raspberry Pi 4B (4GB of RAM). The Raspberry Pi 4B can only manage around 2-3 frames per second. I am considering adding a USB accelerator (like Coral edge TPU processor) to speed it up.

All the procedures are learnt from the official object detection tutorial found in: