Accuracy .
Closed this issue · 1 comments
Hi,
Thanks for your work.
I just gave it a run .. to have a feeling how it is working
the network is not really giving the right object .. it gives a prediction of tv monitor while im showing a bottle.
I tried it with and Android mobile without GPU and another Android with GPU. they needed almost the same time. I just noticed that the GPU Android get heated quickly. maybe capture image and send it for prediction would be good idea instead of keep sending frames.
one last thing, any recommendations for the network size? the pb file? I have around 1800 images and many objects inside them.
Hi,
I've used the tiny-yolov2 model for object detection in this sample application, which is not the most accurate one, however it is small and doesn't require much resources. If you have an android device with higher performance you can try out the yolov2 model, which is deeper and more accurate.
You can find more information about these models on this link: https://pjreddie.com/darknet/yolov2/.
Unfortunately, I did not have a better device, so I built a client-server architecture to achieve better accuracy: https://github.com/szaza/set-detector;
You can find the implementation of an object detection web service here: https://github.com/szaza/tensorflow-java-examples-spring; It also contains a live demo application.