Home

Awesome

Thundernet ncnn

the C++ version of thundernet with ncnn

模型均来自thundernet_mmdetection

ncnn编译对应环境库替换include和lib

检测结果

欢迎各位开发者移植到移动端,并测试贡献耗时

本耗时测试为macos 单线程

MobileNetV2-YOLOv3 (来自ncnn bencnmark)

input shapemAPcost(ms)
352*3520.71567.79

thundernet_shufflenetv2_15_voc

input shapemAPcost(ms)
320*3200.71257.57
352*3520.72264.33
384*3840.73473.63
416*4160.73889.28
448*4480.74497.97
480*4800.747110.04

thundernet_shufflenetv2_15_voc_fpn

input shapemAPcost(ms)
320*3200.7367.71

thundernet_shufflenetv2_15_v2_voc (使用了coco预训练模型)

input shapemAPcost(ms)
320*3200.74964.51
480*4800.778137.49

thundernet_shufflenetv2_15_v2_coco

input shapeAP(0.5:0.95)cost(ms)
320*3200.2272.68
mkdir build
cd build 
cmake ..
make 
./thundernet_voc ../imgs/person.jpg
./thundernet_fpn_voc ../imgs/person.jpg
./thundernet_coco ../imgs/person.jpg

检测结果展示

avatar avatar