【多人姿态估计】AlphaPose推理demo复现

太过爱你忘了你带给我的痛 2022-11-07 04:24 405阅读 0赞

代码参考：[https://github.com/MVIG-SJTU/AlphaPose][https_github.com_MVIG-SJTU_AlphaPose]

文档很齐全，在docs目录下。

AlphaPose做姿态关键点检测，利用yolov3行人检测+姿态关键点检测。

AlphaPose多人姿态估计，利用yolov3行人检测+姿态关键点检测+行人重识别算法。

# 1.环境 #

ubuntu16.04
    cuda9.0
    
    torch==1.1.0
    torchvision==0.3.0
    opencv-python==4.1.0.25
    tqdm
    natsort
    Cython
    pycocotools

python setup.py build develop --user

![20210311115056521.png][]

apt-get update
    apt-get -y install python3.6-tk

# 2.模型准备 #

参考此处：[https://github.com/MVIG-SJTU/AlphaPose/tree/master/trackers][https_github.com_MVIG-SJTU_AlphaPose_tree_master_trackers]

## （1）行人检测模型 ##

[yolov3-spp.weights模型][yolov3-spp.weights]下载，放到AlphaPose/detector/tracker/data/目录下。

![20210311124954355.png][]

## （2）姿态关键点检测模型 ##

[https://github.com/MVIG-SJTU/AlphaPose/blob/master/docs/MODEL\_ZOO.md][https_github.com_MVIG-SJTU_AlphaPose_blob_master_docs_MODEL_ZOO.md]

这里我选择：

<table> 
 <tbody> 
  <tr> 
   <th>Model</th> 
   <th>Backbone</th> 
   <th>Detector</th> 
   <th>Input Size</th> 
   <th>AP</th> 
   <th>Speed</th> 
   <th>Download</th> 
  </tr> 
  <tr> 
   <td><a href="https://github.com/MVIG-SJTU/AlphaPose/blob/master/configs/coco/resnet/256x192_res50_lr1e-3_2x-dcn.yaml" rel="nofollow">Fast Pose (DCN)</a></td> 
   <td>ResNet50 - dcn</td> 
   <td>YOLOv3</td> 
   <td>256x192</td> 
   <td>72.8</td> 
   <td style="width:188px;">2.94 iter/s</td> 
   <td style="width:188px;"><a href="https://drive.google.com/open?id=1zUz9YIk6eALCbZrukxD7kQ554nhi1pVv" rel="nofollow">model</a>&nbsp; 或者 <a href="https://download.csdn.net/download/qq_35975447/15725248" rel="nofollow">我的链接</a></td> 
  </tr> 
 </tbody> 
</table>

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70][]

放到pretrained\_models目录下。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1][]

## （3）行人重识别模型（跟踪） ##

从[这里][Link 1]下载，外网上不去。放到目录下：

trackers/weights/osnet\_ain\_x1\_0\_msmt17\_256x128\_amsgrad\_ep50\_lr0.0015\_coslr\_b64\_fb10\_softmax\_labsmth\_flip\_jitter.pth

# 3.demo可视化测试 #

参考：[https://github.com/MVIG-SJTU/AlphaPose/blob/master/docs/GETTING\_STARTED.md][https_github.com_MVIG-SJTU_AlphaPose_blob_master_docs_GETTING_STARTED.md]  / docs/GETTING\_STARTED.md

python scripts/demo_inference.py --cfg configs/coco/resnet/256x192_res50_lr1e-3_2x-dcn.yaml --checkpoint pretrained_models/fast_dcn_res50_256x192.pth --indir examples/demo/ --vis --showbox --save_img

运行这个，是对examples/demo目录下的几张图片上测试结果。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2][]

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3][]

其中结果保存在examples/res/与examples/res/vis/目录下。

跟踪的话，在上面命令基础上加--pose\_track：

python scripts/demo_inference.py --cfg configs/coco/resnet/256x192_res50_lr1e-3_2x-dcn.yaml --checkpoint pretrained_models/fast_dcn_res50_256x192.pth --indir examples/demo/ --vis --showbox --save_img --pose_track

如果使用了很多图片或者一个视频，需要加上--sp参数。

![watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4][]

![20210311214931351.png][]

[https_github.com_MVIG-SJTU_AlphaPose]: https://github.com/MVIG-SJTU/AlphaPose
[20210311115056521.png]: /images/20221023/62f43b595101412492b78e044d7bf042.png
[https_github.com_MVIG-SJTU_AlphaPose_tree_master_trackers]: https://github.com/MVIG-SJTU/AlphaPose/tree/master/trackers
[yolov3-spp.weights]: https://github.com/Zhongdao/Towards-Realtime-MOT#pretrained-model-and-baseline-models
[20210311124954355.png]: /images/20221023/9fd8f7135c4a4dbbb3a54742d4a5c8ef.png
[https_github.com_MVIG-SJTU_AlphaPose_blob_master_docs_MODEL_ZOO.md]: https://github.com/MVIG-SJTU/AlphaPose/blob/master/docs/MODEL_ZOO.md
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70]: /images/20221023/506e05a904514ea78a49b0c69bd36178.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 1]: /images/20221023/948d2ef603fa42fd9ae8aeebef9f31c7.png
[Link 1]: https://download.csdn.net/download/qq_35975447/15725292
[https_github.com_MVIG-SJTU_AlphaPose_blob_master_docs_GETTING_STARTED.md]: https://github.com/MVIG-SJTU/AlphaPose/blob/master/docs/GETTING_STARTED.md
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 2]: /images/20221023/0811b21538564a28867aaa965c8248a8.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 3]: /images/20221023/e49a927ca38e414b9fa784dc2b0c4329.png
[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzM1OTc1NDQ3_size_16_color_FFFFFF_t_70 4]: /images/20221023/c18ca26370ea43e68951a1fe53551bc4.png
[20210311214931351.png]: /images/20221023/f87975544c084175bd8c77d7eece7ccb.png