|
| 1 | +# YOLOX-Pose |
| 2 | + |
| 3 | +[YOLO-Pose: Enhancing YOLO for Multi Person Pose Estimation Using Object Keypoint Similarity Loss](https://arxiv.org/abs/2204.06806) |
| 4 | + |
| 5 | +## 如何运行 |
| 6 | + |
| 7 | +以下为x86 Linux平台的运行过程 |
| 8 | + |
| 9 | +### 准备 |
| 10 | + |
| 11 | +**python环境** |
| 12 | + |
| 13 | +为了导出模型,您需要安装以下python包 |
| 14 | + |
| 15 | ++ mmengine >= 0.7.1 |
| 16 | ++ mmcv >= 2.0.0rc4 |
| 17 | ++ mmdet >= 3.0.0rc6 |
| 18 | ++ mmyolo >= 0.6.0 |
| 19 | + |
| 20 | +**ncnn** |
| 21 | + |
| 22 | +可以从ncnn的[Release](https://github.com/Tencent/ncnn/releases)页面直接下载预编译包,或者按照ncnn的[wiki](https://github.com/Tencent/ncnn/wiki/how-to-build)从源码进行编译安装 |
| 23 | + |
| 24 | +如果您选择预编译包,解压缩后将ncnn-xxxx-xx目录移动到detective目录下,并重命名为ncnn |
| 25 | + |
| 26 | +如果您从源码编译,将编译后得到的install目录移动到detective目录下,并重命名为ncnn |
| 27 | + |
| 28 | +**opencv-mobile** |
| 29 | + |
| 30 | ++ 如果您已经安装了opencv,可以选择跳过这一步,并相应地修改CMakeLists.txt,使其能够链接opencv |
| 31 | + |
| 32 | ++ 如果您是从源码编译ncnn,并且开启了NCNN_SIMPLEOCV选项(如下所示),同样可以跳过这一步,删除CMakeLists.txt中OpenCV的部分 |
| 33 | + |
| 34 | +```cmake |
| 35 | +option(NCNN_SIMPLEOCV "minimal opencv structure emulation" ON) |
| 36 | +``` |
| 37 | + |
| 38 | ++ 可以从opencv-mobile的[Release](https://github.com/nihui/opencv-mobile/releases)页面选择一个版本下载预编译包,解压缩后移动到detective目录下,并重命名为opencv-mobile |
| 39 | + |
| 40 | +**pnnx** |
| 41 | + |
| 42 | +可以从pnnx的[Release](https://github.com/pnnx/pnnx/releases)页面直接下载预编译包 |
| 43 | + |
| 44 | +**目录结构** |
| 45 | + |
| 46 | +在使用opencv-mobile的情况下,当前工程应当有如下结构 |
| 47 | + |
| 48 | +``` |
| 49 | +detective |
| 50 | +├── assets |
| 51 | +├── yolox-pose |
| 52 | +├── ncnn |
| 53 | +│ ├── bin |
| 54 | +│ ├── include |
| 55 | +│ └── lib |
| 56 | +├── opencv-mobile |
| 57 | +│ ├── bin |
| 58 | +│ ├── include |
| 59 | +│ ├── lib |
| 60 | +│ └── share |
| 61 | +├── ... |
| 62 | +├── LICENSE |
| 63 | +├── README.md |
| 64 | +``` |
| 65 | + |
| 66 | +### 导出 |
| 67 | + |
| 68 | +模型的[结构](https://github.com/open-mmlab/mmyolo/blob/main/configs/yolox/pose/yolox-pose_tiny_8xb32-300e-rtmdet-hyp_coco.py)和[权重](https://download.openmmlab.com/mmyolo/v0/yolox/pose/yolox-pose_s_8xb32-300e-rtmdet-hyp_coco/yolox-pose_s_8xb32-300e-rtmdet-hyp_coco_20230427_005150-e87d843a.pth)文件来自mmyolo的实现。 |
| 69 | + |
| 70 | +YOLOX-Pose的模型中含有Focus层,如果使用pytorch->onnx->ncnn的导出方式,得到的onnx算子比较碎,且需要手工修改ncnn的param文件,比较麻烦,这里我们采用pnnx的导出方式。关于pnnx的详细介绍可参考[pnnx](https://github.com/pnnx/pnnx)。另外,mmdet中Focus实际包含了Focus和Conv两个算子,这里我们将其拆开,使Focus不需要导出权重 |
| 71 | + |
| 72 | +1. 下载[权重](https://download.openmmlab.com/mmyolo/v0/yolox/pose/yolox-pose_s_8xb32-300e-rtmdet-hyp_coco/yolox-pose_s_8xb32-300e-rtmdet-hyp_coco_20230427_005150-e87d843a.pth)文件,并放到export目录下; |
| 73 | +2. 导出pt文件: |
| 74 | +```bash |
| 75 | +cd detective/yolox-pose/export |
| 76 | +python export.py |
| 77 | +``` |
| 78 | +3. 得到model.pt后,导出ncnn |
| 79 | +```shell |
| 80 | +./pnnx model.pt \ |
| 81 | + inputshape=[1,3,640,640] \ |
| 82 | + inputshape2=[1,3,416,416] \ |
| 83 | + moduleop=mmdet.models.backbones.csp_darknet.Focus |
| 84 | +``` |
| 85 | +4. 将转换后的model.ncnn.param和model.ncnn.bin重命名为yolox-pose-tiny.param和yolox-pose-tiny.bin,并放到assets目录下 |
| 86 | + |
| 87 | +### 运行 |
| 88 | + |
| 89 | +```shell |
| 90 | +cd detective/yolox-pose |
| 91 | +mkdir -p build |
| 92 | +cd build/ |
| 93 | +cmake .. |
| 94 | +make -j4 |
| 95 | +./yoloxpose ../../assets/person.jpg |
| 96 | +``` |
| 97 | + |
| 98 | +这里对关键点的可视化参考了[detectron2](https://github.com/facebookresearch/detectron2/blob/main/detectron2/utils/visualizer.py) |
| 99 | + |
| 100 | + |
| 101 | +## 感谢 |
| 102 | + |
| 103 | ++ [ncnn](https://github.com/Tencent/ncnn) |
| 104 | ++ [opencv-mobile](https://github.com/nihui/opencv-mobile) |
| 105 | ++ [mmyolo](https://github.com/open-mmlab/mmyolo) |
| 106 | ++ [detectron2](https://github.com/facebookresearch/detectron2) |
0 commit comments