Which model should I use for object recognition on mobile devices?

Question

I am working on a project where I need the app to recognize a special character, just one small graphic, from a photographed document. Something similar to the example from the picture. More specifically, the app would use this character to determine the corners of the document.

Something like this

Which model would be suitable for that, Mobile SSD, Yolo or something completely different? Approximately how many photos and how much time would it take me to successfully train the model to 90%+ detection? And is TensorFlow Model Maker a good option?

I already tried to train it with Model Maker but the results were really disappointing. I have used

efficientdet_lite0

model. The photos were taken with a phone in high resolution, tagged with labelImg. About 40 for training, five each for validation and test.

It would mean a lot to me if someone would tell me if I am at least on the right track. Thank you very much in advance.

Answer 1

efficientdet or yolo should be good enough for you use case. Yolov5 recommends 10000 annotated instances of each class for good results, this is given that the variability of the class representation is large. I would not start with anything less than a few hundred.