Following the How To Run Android NPU Demo On VIM3 Or VIM3L | Khadas Documentation ，I rebuild the demo code with android studio， then app-debug.apk generated. After adb install -t app-debug.apk, I could see the app on the desk of vim3 phone.
Yet , when I double click the app icon, choosing the any menu such as yolov3 yolov2 etc, nothing changed. My USB UVC camera is ok, we could see the real time images of the real world. The resolution of the camera images is about 600x400.
What was wrong with my settings？ I think there is some possible reason as following.
1 In the project khadas_android_npu_app, what is the meaning of "exec(“chmod 666 /dev/galcore”);” which is the line 314. There is no /dev/galcore directory on my os. And also what is the meaning of “exec(“chmod 777 data”);”?
2 In the line 84 “private static final Size DESIRED_PREVIEW_SIZE = new Size(1920, 1080);” , should I changed the size of image with my camera resolution?
3 Where should I copy the “yolov3_88.nb” “imagenet_slim_labels.txt” and “inceptionv3_88.nb” to seperately?
Looking forwad to your reply!
Note: My galcore version is the 126.96.36.199 , the android image it the latest (about PIE 211220) one.The npu-app is the latest one.
chmod 666 /dev/galcore solve the galcore dev sometimes can not open
chmod 777 data is to add write permisssion when copy nb file
yes ,you can changed the size of image with my camera resolution
yolov3_88.nb is generated by model convert tool
1 In the demo apk like yolov3，how many fps vim3 could get？
2 What is the average and the peak frame per second with these apk demo？
3 How do we eval the time cost of these demo，ang log enable when rebuild or any button useful
4 Also I think it is not directly change the camera resolution ，as the code lined 805 to 809 “imageLocationRectF[index].top = 1080 * (y - height / 2.0f);” ，should be improved too？ Right？
And any other code should be improved？
“3 How do we eval the time cost of these demo，ang log enable when rebuild or any button useful
In fact, I test the yolov3 model in the java code of demo apk, the cost time of npu detection is almost more than 100ms every frame. It is too slow with respect to NPU. I want to know the yolov3 nb model here is fp32 or fp16, why is it so slow ? If the nb model is int8 quant version, it will be more effecient than this! Could you have a deep look at this ?
Could you have a look my question？ tks
At present, npu app is only a simple demo, and its performance needs to be optimized
1 Yet，with respect the latest Tegnine，could you please shared the vim3 npu android sdk and related lib to me ？
2 Any plan for the latest pose model net，such as tinypose？