Ok, so I have converted a custom Tensorflow segmentation model to DLC and tested it on ARM.
When running the model on some video frames, via executable "snpe-sample" in ADB , the average runtime is 20ms (GPU).
I have also modified the Android tutorial application to perform a segmentation task on the selected image, and the runtime is about 1500ms (GPU).
What is the reason of this x75 increase in time? Is there any way of bulding a reliable SNPE Android application that performs as in benchmarks/direct ADB executable run?
Best regards,
Calin