WIP: Tinymaix

[bkmgit]

Example showing how to use CMake to build a library.

The TinyMaix library can probably be integrated, though the bindings may need to not use code derived from TinyMaix to be incorporated.

The current example uses a network trained on the MNIST dataset - licensing of this is unclear, though it is an open dataset. This likely cannot be integrated. Perhaps train a network on the Dry Bean dataset?

[bkmgit]

Builds and runs, but recognition rate is poor, see sipeed/TinyMaix#93 . May need to use alternate model.

[bkmgit]

@mgorges @ddunsmuir Would appreciate an initial review. There is an extended MNIST dataset which one could use to train a more accurate model. The present model works when one takes a picture of a single number on a white background. The number should be centered in the image. One could use a Maximally Stable Extremal Regions to extract the location of the digit, either programming it directly or using a library such as OpenCVMobile or VLfeat, at the moment cropping is used to extract a square region in the middle of the image.

[AgoeZero]

During my initial testing on Debian Trixie, I found an issue when TinyMaix is installed for the first time. The CMake build attempted to invoke ranlib using a relative path instead of the absolute path provided by the toolchain, causing the build to fail. After hardcoding the correct ranlib path, the library built successfully. This only appears to affect the initial installation, since subsequent builds reuse the installed library. I'm continuing to test the TinyMaix examples and will report any additional observations.

[bkmgit]

Thanks. Do you have a patch indicating what you hard coded? Path to ranlib is set here, do builds for packages which use configure work for you without changes? Path to ranlib for configure is set here