the C++ version of thundernet with ncnn
模型均来自thundernet_mmdetection
在ncnn编译对应环境库替换include和lib
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 shape |
mAP |
cost(ms) |
352*352 |
0.715 |
67.79 |
thundernet_shufflenetv2_15_voc
input shape |
mAP |
cost(ms) |
320*320 |
0.712 |
57.57 |
352*352 |
0.722 |
64.33 |
384*384 |
0.734 |
73.63 |
416*416 |
0.738 |
89.28 |
448*448 |
0.744 |
97.97 |
480*480 |
0.747 |
110.04 |
thundernet_shufflenetv2_15_voc_fpn
input shape |
mAP |
cost(ms) |
320*320 |
0.73 |
67.71 |
thundernet_shufflenetv2_15_v2_voc (使用了coco预训练模型)
input shape |
mAP |
cost(ms) |
320*320 |
0.749 |
64.51 |
480*480 |
0.778 |
137.49 |
thundernet_shufflenetv2_15_v2_coco
input shape |
AP(0.5:0.95) |
cost(ms) |
320*320 |
0.22 |
72.68 |
mkdir build
cd build
cmake ..
make
./thundernet_voc ../imgs/person.jpg
./thundernet_fpn_voc ../imgs/person.jpg
./thundernet_coco ../imgs/person.jpg