FaceBoxes is a cpu real-time face detector with high accuracy

ncnn is a high-performance neural network inference computing framework optimized for mobile platforms.

how to train

repair some bug in onnx2ncnn

1.onnx2ncnn in ncnn don't support the operation which changes Split to Slice. Modifying the onnx2ncnn on the basis of ncnn in order to support Split to Slice operator.

2.onnx2ncnn in ncnn would parse to the wrong parameter in reshape operator if reshape operator is 4 dims(eg:torch.reshape(b, c, h, w) to ncnn reshape).

3.winograd3x3 conv style has bug, I don't use winograd3x3.

1. set opencv path in ./faceboxes/CMakeLists.txt file

2. how to build ncnn library

3. cd ./build/faceboxes/ && ./facebox, test.jpg saved in ./model/test.jpg

We provide the converted model in model folder!

1.ncnn

2.facebox