Re-train a cnn using transfer learning and a custom rgb dataset and deploy on Jetson Nano

I have a rgb dataset of 10500 images (224x224). I want to implement an object detection system in my Jetson Nano 4gb. You can use the popular docker container which contains SSD Mobilenet v2, you should choose a cnn and retrain it with my dataset.

AC: In order to accept the project I would like to test in real time the model (the dataset is about sign language, so I should make a sign in front of my camera and the system should show the label of the such sign)

There is not neccesary that you re-train the cnn with the whole dataset, you can take just a portion (but all classes included). If the systems achieves a good accuracy with less data... cool.