Skip to content

Latest commit

 

History

History
87 lines (71 loc) · 3.88 KB

readme.md

File metadata and controls

87 lines (71 loc) · 3.88 KB

功能

  • 基于c++14实现top-down multi-person whole body estimation
  • 目标检测yolox-s ncnn
  • 姿态估计fastpose 基于libtorch或ncnn
  • vs2019开发
  • 模型文件仓库

package版本


# onnx 转 ncnn
```shel
.\onnx2ncnn.exe models\halpe26_fast_res50_256x192.onnx models\halpe26_fast_res50_256x192.param models\halpe26_fast_res50_256x192.bin
.\ncnnoptimize.exe models\multi_domain_fast50_regression_256x192.param models\multi_domain_fast50_regression_256x192.bin models\multi_domain_fast50_regression_256x192.opt.param models\multi_domain_fast50_regression_256x192.opt.bin 65536

run cli

  • fp32
.\AlphaPose.exe -dpm ..\..\AlphaPose\models\yolox_s.opt.param -dbm ..\..\AlphaPose\models\yolox_s.opt.bin -ppm ..\..\AlphaPose\models\multi_domain_fast50_regression_256x192.opt.param -pbm ..\..\AlphaPose\models\multi_domain_fast50_regression_256x192.opt.bin -pj 136 -i .\1.jpg -o 1_out.jpg  -dt 4 -pt 4 -wc 5
.\AlphaPose.exe -dpm ..\..\AlphaPose\models\yolox_s.opt.param -dbm ..\..\AlphaPose\models\yolox_s.opt.bin -ppm ..\..\AlphaPose\models\halpe26_fast_res50_256x192.opt.param -pbm ..\..\AlphaPose\models\halpe26_fast_res50_256x192.opt.bin -pj 26 -i .\1.jpg -o 1_out.jpg  -dt 4 -pt 4 -wc 5
 -dpm  ..\alpha_pose_models\yolox_nano.opt.param -dbm ..\alpha_pose_models\yolox_nano.opt.bin -ppm ..\alpha_pose_models\halpe26_fast_res50_256x192.opt.param -pbm ..\alpha_pose_models\halpe26_fast_res50_256x192.opt.bin  -pj 26  -i pics\1.jpg -o pics\1-out.jpg -dt 4 -pt 4 -wc 5
  • fp16
 -dpm  ..\alpha_pose_models\yolox_nano-opt.param -dbm ..\alpha_pose_models\yolox_nano-opt.bin -ppm ..\alpha_pose_models\halpe26_fast_res50_256x192-opt.param -pbm ..\alpha_pose_models\halpe26_fast_res50_256x192-opt.bin  -pj 26  -i pics\1.jpg -o pics\1-out.jpg -dt 4 -pt 4 -wc 5 -fp16

yolov5lite-e fp16

-dpm  ..\alpha_pose_models\yolov5lites_fp16\yolov5lite-e.param -dbm ..\alpha_pose_models\yolov5lites_fp16\yolov5lite-e.bin -ppm ..\alpha_pose_models\halpe26_fast_res50_256x192-opt.param -pbm ..\alpha_pose_models\halpe26_fast_res50_256x192-opt.bin  -pj 26  -i pics\1.jpg -o pics\1-out.jpg -dt 4 -pt 4 -wc 5 -fp16

time cost

Object Detector input height: 416, input width: 416
AlphaPose model load and init time: 1.07343
Yolox trans mat time: 0.0039237
Yolox infer mat time: 0.0382188
Yolox gen box time: 0.0001189
Yolox nms box time: 1.14e-05
Object detected time: 0.0454368
Preprocessed image tensor 0 time: 0.0009864
ncnn fastpose forward 0 time: 0.0600666
ncnn mat convert torch tensor0 time: 0.0003712
Generate landmarks 0 time: 0.0125228
Pose detected time: 0.0765909
AlphaPose model infer time: 0.122566

models

  • fp32
yolox_nano.opt
halpe26_fast_res50_256x192
halpe26_fast_res50_256x192-fp32
  • fp16
yolox_nano-opt
halpe26_fast_res50_256x192-opt

convert fp16 int8

  • int8
/ncnn2table yolov5-lite-opt.param yolov5-lite-opt.bin imagelist.txt yolov5-lite.table mean=[104,117,123] norm=[0.229f, 0.224, 0.225] shape=[640,640,3] pixel=BGR thread=8 method=kl

run

-dpm  ..\alpha_pose_models\yolov5lites_fp16\yolov5lite-e.param -dbm ..\alpha_pose_models\yolov5lites_fp16\yolov5lite-e.bin -pm ..\alpha_pose_models\rtmpose-s-halpe  -pj 136  -i pics\1.jpg -o pics\1-out.jpg -dt 4 -pt 4 -wc 5  -id 1