fatal error: opencv2/dnn/dnn.hpp: No such file or directory

[andregirard37]

Hello,

I am building a Yocto image with Dunfell but adding in my custom layer the opencv_4.5.5 recipe as well as the bbappend for CUDA dnn, that you could find on the Kirkstone release.
After building the rootfs, I believe I have cudnn available. However, for development purposes, I require the sysroot I am generating to have the dnn header files (ex: opencv2/dnn/dnn.hpp), for cross compiling applications on my host laptop. I tried adding as many opencv* recipes I could but unsuccessful. I added:
opencv
opencv-samples
libopencv-core-dev
libopencv-highgui-dev
libopencv-imgproc-dev
libopencv-ml-dev
Any idea what recipe I need to add that would include all the dnn header files ?
Thanks!

[dwalkes]

However, for development purposes, I require the sysroot I am generating to have the dnn header files (ex: opencv2/dnn/dnn.hpp), for cross compiling applications on my host laptop.

Are you using the Extensible SDK or Standard SDK for your host laptop development? I would expect header files to end up in the SDK but not in the target rootfs. I suspect there's probably not a recipe for including development header files on the target rootfs today.

[andregirard37]

I am a bit confused on the following line in the recipe
PACKAGECONFIG[dnn] = "-DBUILD_opencv_dnn=ON -DPROTOBUF_UPDATE_FILES=ON -DBUILD_PROTOBUF=OFF,-DBUILD_opencv_dnn=OFF,protobuf protobuf-native,"
why is there DBUILD_opencv_dnn ON then OFF ? Is that to list the 2 available options ?
And also in the bbappend:
-DWITH_CUDA=OFF
Shouldn't that be ON
Thanks

[andregirard37]

in to log.do_configure I see:
-- OpenCV modules:
-- To be built: alphamat aruco bgsegm bioinspired calib3d ccalib core cudaarithm cudabgsegm cudacodec cudafeatures2d cudafilters cudaimgproc cudalegacy cudaoptflow cudastereo cudawarping cudev datasets features2d flann fuzzy gapi hfs highgui img_hash imgcodecs imgproc intensity_transform line_descriptor ml optflow phase_unwrapping photo plot python3 quality rapid reg rgbd saliency sfm shape stereo stitching structured_light superres surface_matching tracking ts video videoio videostab xfeatures2d ximgproc xphoto
-- Disabled: dnn freetype text world
-- Disabled by dependency: barcode cudaobjdetect dnn_objdetect dnn_superres dpm face mcc objdetect wechat_qrcode xobjdetect
-- Unavailable: cvv hdf java julia matlab ovis python2 viz
-- Applications: perf_tests examples apps
-- Documentation: NO
-- Non-free algorithms: NO

So there might be one recipe/dependency that is preventing dnn to build correctly ?

[madisongh]

Yocto documentation on PACKAGECONFIG

The Disabled: line means that OpenCV's dnn module was disabled in the configuration. So something must not be right with the back-port of your bbappend, which should be enabling cuda and dnn in the configuration automatically for the Tegra platforms.

[dwalkes]

I think the problem is that the TEGRA_CUDA_ARCHITECTURE check here doesn't work when copying to the dunfell branch because the commit which added this at 4a89e2b isn't used in the dunfell tegra-194.inc.

I've verified if I add TEGRA_CUDA_ARCHITECTURE = "72" to tegra194.inc I get:

-- Found CUDNN: /home/dan/trellis/tegra-demo-distro/build-dunfell/tmp/work/armv8a_tegra194-oe4t-linux/opencv/4.5.5-r0/recipe-sysroot/usr/lib/libcudnn.so (found suitable version "8.2.1", minimum required is "7.5")

and

--   cuDNN:                         YES (ver 8.2.1)

in log.do_configure.

[andregirard37]

Thanks a lot! I added in my local.conf :

TEGRA_CUDA_ARCHITECTURE = "72"

and now I see the same lines:

-- Found CUDNN: /home/yocto/projects/l4w-nx/build/tmp/work/armv8a_tegra194-poky-linux/opencv/4.5.5-r0/recipe-sysroot/usr/lib/libcudnn.so (found suitable version "8.2.1", minimum required is "7.5")

And:

--   cuDNN:                         YES (ver 8.2.1)

Now I see the dnn/ folder in my sysroot! Thanks for all the support. Amazing!