I have implemented the LFFD referring to the official python implementation the Inference time of LFFD with the input shape of 320x240 is about 20ms on the Qualcomm Snapdragon 632 CPU
paper:LFFD: A Light and Fast Face Detector for Edge Devices
official github: LFFD
My MNN implementation MNN.
My OpenVINO implementation
- You can set the input tensor shape smaller ,since you need to reduce the memory and accelerate the inference.
- You can set the scale_num=8 to use another larger model.
- I just test it on vs2019 PC and the result is correct compared to original implementation,you can use the code to another device such as android、RK3399、and so on.
The original mxnet model has merged the preporcess(means and norms) and the detection output tensor has been sliced with the mxnet slice op in the symbol ,which caused convert failure. so,you need to remove these ops ,in that way you can convert the model to onnx/ncnn successfully.I will show you how to do that step by step, so when you train the model by yourself, you can convert to your own model to onnx , and do more things.
-
First ,follow the author's original github to build the devolopment environment.
-
Modify symbol_10_320_20L_5scales_v2.py (your_path/A-Light-and-Fast-Face-Detector-for-Edge-Devices\face_detection\symbol_farm)
in function loss_branch,Note out（注释掉） the line 57(predict_score = mxnet.symbol.slice_axis(predict_score, axis=1, begin=0, end=1)
in function get_net_symbol, Note out（注释掉）the line 99(data = (data - 127.5) / 127.5,preprocess).
-
Next,in this path , by doing "python symbol_10_320_20L_5scales_v2.py ",generate the symbol.json. symbol_10_560_25L_8scales_v1.py do the same thing .
- MNN demo finished
- openvino demo: mxnet model-->onnx-->openvino
- TensorRT demo: mxnet model --> onnx-->trt engine(coming soon)