How to run the face detection/recognition on edge tpu device ? about smart-zoneminder HOT 9 CLOSED

Hi,

I've gotten face detection and recognition to work on Google Coral using tpu-, dlib- and openCV-based face detectors and dlib- and OpenCV-based facial embedders and an SVM-based face classifier. The detailed instructions on how to set this up can be found here (note that you don't need to follow all the steps outlined here if you are just interested in face detection / recognition). Please take a look at that and let me know if you have any questions.

Thanks.

Lindo

from smart-zoneminder.

Closing due to inactivity.

from smart-zoneminder.

Hi,
Sorry for the inactivity. Could you please reopen the issue ?
Thanks,
Vismaya

from smart-zoneminder.

As you wish. Please let me know how I can help.

from smart-zoneminder.

Hi,
Thanks a lot. Is it possible to do the face recognition module on edge tpu device without memory swap ? When I tried it last time, the system got crashed.

from smart-zoneminder.

Do you mean the temporary swap space required to compile dlib and face_recognition on the coral board? Yes, you can cross compile those packages on another machine instead. But not sure what your problem is exactly. The swap space is temporary and after you are done compiling the packages you remove the swap. See the detailed instructions at https://github.com/goruck/edge-tpu-servers/blob/master/README.md. You can send me the commands you are using and the error and I will try to help further.

from smart-zoneminder.

I just tried swaping for the installation of OpenCV, the board crashed and I lost all the data on the board. But I've compiled OpenCV now with swaping. So my question here is can we compile dlib and face_recognition without generating swap space. I've generated all the models including svm and xgboost. How do I test the models on desktop ? Only "encodings.pickle" is working. I'm not able to test the svm and xgboost models. If everything goes well and fine how much time will it take for inference on the board ? and how do I find the accuracy and loss ?

from smart-zoneminder.

The Coral Dev Board has a modest 8GB on-board eMMC. You need to insert a MicroSD card (at least 32 GB) into the Dev Board to have enough space to install the software. Are you using a MicroSD card?

As for your questions:

1. "can we compile dlib and face_recognition without generating swap space"
   No. You will run out of ram and the compiler will crash.

2. "How do I test the models on desktop?"
   See face-det-rec and in particular train.py.

3. "If everything goes well and fine how much time will it take for inference on the board" + "how do I find the accuracy and loss".
   See this.

from smart-zoneminder.

Thanks for the help 👍

from smart-zoneminder.