juzisyj/centerface-ncnn

using ncnn to deploy centerface on Android

centerface-ncnn

using ncnn, on android,detection,surport arm v7/v8/x86, surport vulkan on gpu

environment

ncnn
AndroidStudio3.2(NDK19)
Qualcomm845

model

using pretrained centerface ,bystar-clouds
pipeline: model on PC --> ONNX ---> model on ncnn(.bin,,param) --> inference on Android

cost_time(only once)

item	PC(i7 8700 500*300)	Android 模拟 600*400
load_model	5ms	35ms
infer_time	192ms	190ms
decode_time	0ms	1ms

organization

using ncnn and opencv on android
the sim-overall android-organization is here

result

infer time on 845 in float32 is aroud 350ms(700x500), depending on the object

ncnn is good in optimazing on float32

reference

(https://github.com/JuZiSYJ/MobilenetSSD_Android)

(https://github.com/nihui/ncnn-android-squeezenet)

(https://blog.csdn.net/qq_33431368/article/details/84990390)