bits and pieces for detecting North American-style (MUTCD) speed limit signs in images
So far there isn't too much here:
train_signmatch.py - train the object detector using photos with signs
in them
signmatch.py - use the training data to find signs in other photos
speedlimits.svm - training data derived from a corpus of sign photos
To use this, you'll need to build
dlib with Python support. You
will need Python 2.7, cmake, libjpeg, boost-python, and
scikit-image (aka python-skimage). The code will probably work with
Python 3.x as well, but is as-yet untested.
To build the Python dlib module,
sh compile_dlib_python_module.bat under the python_examples
directory. Then copy dlib.so into your Python path or this directory
so the Python scripts can use it.
If you want to train the classifier with your own signs, follow the
instructions in Ian Dees' Gist listed below through step 5. Then, use
the train_signmatch.py script (for some reason, the classifiers from
DLib are incompatible between C++ and Python).
On x86, you probably will want to enable AVX. This should be the
default in dlib after version 18.17; until then, you'll need to do
this manually by editing compile_dlib_python_module.bat to add
-DUSE_AVX_INSTRUCTIONS=ON to the first cmake command so it reads:
cmake ../../tools/python -DUSE_AVX_INSTRUCTIONS=ON
All this depends on the awesome DLib project at https://github.com/davisking/dlib/
Idea shamelessly stolen from Ian Dees' Gist at https://gist.github.com/iandees/f773749c47d088705199
Photos used for training were mostly sourced from Wikimedia projects; the rest are my own.