Install object_detect_tf post processing stage #1214
Conversation
|
tensorflow-lite from meta-neural-network currently only builds for 64-bit systems. Please let me know what I must add in a new commit to skip the build of tensorflow-lite for 32-bit systems (libcamera-apps will build just fine if tensorflow was requested by the user but not available in sysroot.) |
| @@ -0,0 +1,4 @@ | |||
| do_install:append () { | |||
There was a problem hiding this comment.
I still find this an issue that needs to be fixed in the actual layer. I'm not sure I understand the reasoning not to do that.
There was a problem hiding this comment.
right this is a generic change.
There was a problem hiding this comment.
Obviously that is a generic change that is being discussed upstream. The reasoning behind having a bbappend with the fix in our layer is to document the issue in the meantime, while still providing a functional example to the users. I'm not sure if you noticed, but processing with tensorflow did not work until now.
There was a problem hiding this comment.
I see that the review is in progress, and it introduces some compile issues. Let's see how that settles upstream (see the symlink discussion). I want to avoid:
- introducing new issues while fixing this
- making sure that this gets cleaned up after upstream deals with it
The libcamera-apps project provides assets for various built-in, OpenCV or TensforFlow Lite post processing stages. Install them as part of the package. Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
Install the object_detect_tf stage described in the official
documentation:
https://www.raspberrypi.com/documentation/computers/camera_software.html#object_detect_tf-stage
With libcamera-apps compiled with tflite support, one is able to run
the object detection model with e.g.:
libcamera-still --post-process-file /usr/share/libcamera-apps/assets/object_detect_tf.json --lores-width 400 --lores-height 300
Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
The tensorflow-lite_2.* recipe from meta-neural-network installs the library and pkg-config files as tensorflow2-lite*, because their users may use several versions of the same library. However packages such as libcamera-apps expect tensorflow-lite* files. While this is being discussed with the upstream meta-neural-network layer, adding a dynamic bbappend in meta-raspberrypi also helps documenting the support for this tensorflow-lite provider. Refs nnstreamer/meta-neural-network#85 Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
The meta-neural-network provider for tensorflow-lite only builds on 64-bit systems. Set COMPATIBLE_MACHINE to skip the build on 32-bit variants of the Raspberry Pi SoC family. Signed-off-by: Vivien Didelot <vivien.didelot@gmail.com>
vivien
force-pushed
the
libcamera-assets
branch
from
ae60230
to
29da28b
Compare
|
I've documented the upstream issue in the code and commit message and skipped the build on 32-bit systems (libcamera-apps builds fine when tflite isn't available in sysroot.) |
With this PR, one can add the meta-neural-network layer to their system, enable tensorflow support in libcamera-apps with e.g.:
and try object detection with the new libcamera-detect app or even
libcamera-still --post-process-file /usr/share/libcamera-apps/assets/object_detect_tf.json --lores-width 400 --lores-height 300for example.(tested on raspberrypi4-64 with a Camera Module 3)