TFDetect

Provides a C++ library to perform inference on models trained with the tensorflow object detection scripts.

Prequesites

• OpenCV 2.4 (packages: libopencv-dev)
• Build tools (packages: build-essentials, autoconf)
• Bazel (see here)

Setup

mkdir build && cd build
cmake -DCMAKE_INSTALL_PREFIX=<FILL_IN> ..
# bazel opts are needed for bazel 0.9
ADDITIONAL_BAZEL_OPTS='--incompatible_load_argument_is_label=false' TF_NEED_MKL=1 make
# you can also build with CUDA support. Note that you will be prompted for CUDA specific configurations during the build
ADDITIONAL_BAZEL_OPTS='--incompatible_load_argument_is_label=false' TF_NEED_MKL=1 TF_NEED_CUDA=1 make
make install

Why Tensorflow C-API?

One might ask himself, why I'm reinventing the wheel by writing a C++ wrapper for the Tensorflow C-API. That's because – unlike the C-API – the tensorflow C++ API is a monster which is linked against many 3rd party libraries. In my case, I had a version clash with the OpenCV library when linking TFDetect to an application. So instead of solving this issue, I simply wrote the custom wrapper.