I suggest you to visit the website of TensorFlow , we can find you the example to explained you how devise your datasets and how you can test your model by using of images tests one by one.

I hope that be Claire for you.@ Nima Karimi

You can use your data for training and testing purposes if you are testing the performance of your method, however., in object detection task the custom way is to use or split you data into 3 parts, Training, test and validation. This is because you have to compare performance of your method to that of the existing approaches that may be based on training, testing, validity.

According to my knowledge, there is no 'specific proportion' to choose; only rules of thumb, such as 80:20 (training vs test). I don't think the proportion that you mentioned as ideal: '70:15:15' is something to be strictly followed. As I have come across rule of thumb of 80: 20 (in literature); I think not including 15% data for validation can be acceptable.

I can suggest the initial chapters of this book as a useful reading relevant to your question:

https://www.oreilly.com/library/view/hands-on-machine-learning/9781492032632/

Three images may not be sufficient to arrive at a conclusion that will be universally applicable. Usually a minimum number of 30 images should be considered for a meaningful conclusion, divided in the ratio 70:15:15. Without a validation set, there might be overfitting in the training set which will lead to a reduced performance of the system.

The standard practice is to have more training data than test data. The proportions recommended are 60:40, 70:30 or 80:20.

If you are going for validation then the ration may be different like 80:10:10.

70:30 is the most recommended one for image data.

I suggest you to visit the website of TensorFlow , we can find you the example to explained you how devise your datasets and how you can test your model by using of images tests one by one.

I hope that be Claire for you.@ Nima Karimi