This is unofficial implementation (inference only) of FaceBoxes: A CPU Real-time Face Detector with High Accuracy, in libtorch (pythorch C++ API). The original caffe implementation can be found here

Move anchor.pt and face.pt from this repo to /usr/models/

• Download libtorch from pytroch.org extract and put it at /usr/lib/
• cd faceboxes_libtorch/ && mkdir build
• cd build && cmake -DCMAKE_PREFIX_PATH=/usr/lib/libtorch ..
• make -j$(nproc)
• put sample.jpg in build folder
• run ./faceboxes result will be saved in build/result.jpg

Tested on Ubuntu 20.04, libtorch 1.7.1 the weight is traced from the weight published here https://github.com/zisianw/FaceBoxes.PyTorch. So the performance is expected to be the same

• Deploy on android
• Use half-precision model
• Libtorch