On our journey to teach a drone to move around a room, we taught it to go around and look around, collecting images.
We then collected a few images and used them to instruct a model with TensorFlow. Our model can distinguish between images where a ball is present (ok) and images where it is not present (ko).
In our first experiment we collected only about 200 images and used a standard convolutional model, shown in the figure.
The first attempt produces limited results but not completely discouraging: 75% accuracy in identifying the ball, but you can certainly do better. We will have to work on improving the dataset.
The notebook we used it is here.