Thank you Alex! According to your reply, Should I use the gpu for large tasks such as Neural Network or large image instead of the task that need to process each frames?

For any of our projects, there should be a LICENSE file in the root directory that dictates the terms of use. Most of our software has a license that limits to use with ModalAI hardware. There are a few exceptions, like libmodal_pipe, which have different licenses for compatibility purposes. You can see the license for libmodal_pipe here: https://gitlab.com/voxl-public/voxl-sdk/core-libs/libmodal-pipe/-/blob/master/LICENSE

@Matt-Turi Ah ok that sounds interesting might be worth a try. Good to know that. Thanks for the fast answer as always.

Solution here

Hey @Philemon-Benner, As of a few minutes ago, the latest voxl-tflite-server (v0.3.0) is now running with tensorflow v2.8.0 for apq8096. This means that you can now run any of the models that were included with qrb5165 on VOXL without dealing with these headaches. You can pull this from the dev apq8096 packages repo, or directly from here: http://voxl-packages.modalai.com/dists/apq8096/dev/binary-arm64/voxl-tflite-server_0.3.0_202207212215.ipk

I realize I made a mistake and meant 12S, but same problem

@Philemon-Benner We build this as part of our Yocto system image build. It is 32 bit. We have not tried to build this outside of the Yocto build.

Thank you for the Answer

@Alex-Gardner Thank you for the fast answer

@Matt-Turi Ok nice i've send an email to you

Hello @Philemon-Benner , It is possible to use any MPA image pipe as input for voxl-tflite-server. For a usb camera, you must be using the latest voxl-uvc-server to get the camera frames from your usb device into our MPA pipes. Then, you must modify the config file (under /etc/modalai/voxl-tflite-server.conf) such that the input pipe matches the new usb pipe: voxl:/$ vi /etc/modalai/voxl-tflite-server.conf /** * This file contains configuration that's specific to voxl-tflite-server. * * skip_n_frames - how many frames to skip between processed frames. For 30hz * input frame rate, we recommend skipping 5 frame resulting * in 5hz model output. * model - which model to use. Currently support mobilenet for * object detection or midas for monocular depth * input_pipe - which camera to use (tracking or hires). */ { "skip_n_frames": 0, "model": "/usr/bin/dnn/ssdlite_mobilenet_v2_coco.tflite", "input_pipe": "/run/mpa/USB_PIPE_NAME/", // CHANGE HERE "delegate": "gpu" } The only limitation with the tflite server is supported image formats. In this code block, there is a switch for the supported formats, so if your input data is not in NV12, NV21, or RAW8 format you will have to add handling for your specific type.

@Chad-Sweet Awesome, thanks!

@Eric-Katzfey microhard addition

Hi Prasad, I've seen the error that you talked about with the ungenerated once or twice and running the calibration procedure again properly generated the files. This is a sample stereo cal from my board (that has a few additional data fields from a soon to be released cal rework) that would serve as a good general idea of what the files should look like. The DFS algorithm is, however, very sensitive to inaccuracies and I'd strongly recommend trying to calibrate your cameras again, but if that doesn't work you can add your values to this format: yocto:/$ cat /data/modalai/opencv_stereo_extrinsics.yml %YAML:1.0 --- R: !!opencv-matrix rows: 3 cols: 3 dt: d data: [ 9.9965050025202740e-01, -5.6519379540328910e-04, 2.6430245966411248e-02, 2.8842568739021546e-03, 9.9613506820545850e-01, -8.7787282413751958e-02, -2.6278478041102546e-02, 8.7832832399280705e-02, 9.9578854941426354e-01 ] T: !!opencv-matrix rows: 3 cols: 1 dt: d data: [ -7.9514399346537298e-02, -2.7245018421472277e-03, 6.0288890076187220e-04 ] reprojection_error: 2.8997817600221831e-01 calibration_time: "2021-09-30 22:52:56" yocto:/$ cat /data/modalai/opencv_stereo_intrinsics.yml %YAML:1.0 --- M1: !!opencv-matrix rows: 3 cols: 3 dt: d data: [ 4.9291881707143079e+02, 0., 3.0643471662712096e+02, 0., 4.9232152259255315e+02, 2.1697996372929907e+02, 0., 0., 1. ] D1: !!opencv-matrix rows: 5 cols: 1 dt: d data: [ -1.6984598364049669e-01, 6.0034285910162230e-02, -8.8644933037586781e-04, -4.1419313003228072e-04, 0. ] reprojection_error1: 1.9741562557305206e-01 M2: !!opencv-matrix rows: 3 cols: 3 dt: d data: [ 4.9673975454370270e+02, 0., 3.1711708934220133e+02, 0., 4.9610254255552854e+02, 2.2876245199481119e+02, 0., 0., 1. ] D2: !!opencv-matrix rows: 5 cols: 1 dt: d data: [ -1.6758727319959599e-01, 5.4041564358266361e-02, -1.5695644102222835e-03, -5.8891426327232642e-04, 0. ] reprojection_error2: 3.1594418176162986e-01 width: 640 height: 480 distortion_model: plumb_bob calibration_time: "2021-09-30 22:52:56"

Thanks for your help. I realized that I hadn't configured the cameras correctly. I ran the voxl-configure-cameras and set it up correctly. Now I am getting the position estimates in QGC.

Hey @colerose, Glad to hear you got it working, and I apologize for the typos in my code snippets. I do not know of any way to send coordinates over rtsp, but if you are willing to work in c++, you are able to access that data directly in the voxl-tflite-server.

I tried everything there was in the tutorial and spent like 1h with different timing. I was able to unbrick other Voxls before but this one doesn't want to. If there is a different method than what you already described please let me know.