Hello, I’m making android app for logo detection.
My first try is yolov3-416 but detection speed is too slow(vsi_nn_RunGraph consumes 150ms at 608x608, 90ms at 416x416 ), so I tried yolov3-tiny but problem occured. vnn_CreateYoloTiny() successed, but vsi_nn_verifyGraph fail.
Is this DDK version problem? I saw your updated aml_npu_app(yolov3-tiny demo added) and DDK version of that is 6.3.3.4. Khadas Android repo’s NPU DDK version is 6.3.2.5(/vendor/amlogic/common/npu). If cause is that, Can I get updated version of NPU DDK so files or repo of Android? Or another cause?
ADD : This error caused by different output tensor size. I fixed and RunGraph time is 10ms.
I made some jni code using NPU so library and build it by ndk-build, and load library in android app.
But I think there is a way porting Android.Neuralnetworks api, not using jni.
@bingbingo
I’m korean and I have no chinese language skill sorry.
Running NPU in Android is very annoying job. You have to fixed some selinux permission for using galcore char device and libovxlib. NPU App needs permission for using this resource. So, you have to build android AOSP. selinux permission file located this directory.
(AOSP source directory)/common/sepolicy/
I decided make my app to system app and give permission to system apps.