Yolopose practice: one-stage human posture estimation with hands + code interpretation

Open source address : https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose

Reading guide ：
I saw a piece of news not long ago ,YOLO The father of Joseph Redmon Leave CV world , The reason is that I can't stand moral condemnation , This technology has been used in military and privacy issues . lately ,YOLO Fire again ,YOLOv7 It has reached the best level in the balance of speed and accuracy . And based on YOLOv5 Of YOLOPose It has also achieved end-to-end leading performance in the field of human posture estimation . This record is reproduced YOLOPose The process of , And code interpretation .

---

---

One 、 Set up

1.1 Cloning of warehouse , Install dependency library , Check Pytorch and GPU.

!git clone https://github.com/ultralytics/yolov5  # clone repo
#YOLO-Pose Branches cannot be cloned , Direct download zip Bag can ;
%cd edge-yolov5-yolo-pose/
%pip install -r requirements.txt  # install dependencies

import torch
from IPython.display import Image, clear_output  # to display images

clear_output()
print(f"Setup complete. Using torch {
      torch.__version__} ({
      torch.cuda.get_device_properties(0).name if torch.cuda.is_available() else 'CPU'})")

Setup complete. Using torch 1.9.1+cu101 (Tesla V100-SXM2-16GB)

If pytorch Wrong version , The following problems occur ：

AttributeError: ‘Upsample‘ object has no attribute ‘recompute_scale_factor‘
#https://blog.csdn.net/qq_40280673/article/details/125095353

# The right version 
pip install torch==1.9.1
pip install torchvision==0.10.1

Local conda Operation in environment ,(pytorch-cifar) [email protected]:edgeai-yolov5-yolo-pose$ pip install torch==1.9.1

Two 、 Reasoning

2. 1 Download the trained YOLO and YOLOPose Model ;

(pytorch-cifar) [email protected]:edgeai-yolov5-yolo-pose$ ./weights/download_weights.sh
Because directly open the link to download , There will be format errors , It is recommended to download from the command line ：

wget http://software-dl.ti.com/jacinto7/esd/modelzoo/gplv3/08_02_00_11/edgeai-yolov5/pretrained_models/checkpoints/keypoint/coco/edgeai-yolov5/other/best_models/yolov5s6_640_57p5_84p3_kpts_head_6x_dwconv_3x3_lr_0p01/weights/last.pt

# Download here Yolov5s6_pose_960

2.2 Prepare the data

Data sets need to be prepared as YOLO Format , Because data loading requires reading the information of key points and location boxes at the same time . There is a special code base that can produce the required format , see https://github.com/ultralytics/JSON2YOLO . Because the author has put coco keypoints Convert to the desired format , Download the attachment package directly . And soft connect the data set to the required folder , The folder directory is as follows ：

edgeai-yolov5
│   README.md
│   ...   
│
coco_kpts
│   images
│   annotations
|   labels
│   └─────train2017
│       │       └───
|       |       └───
|       |       '
|       |       .
│       └─val2017
|               └───
|               └───
|               .
|               .
|    train2017.txt
|    val2017.txt

Refer to the following for soft connection ：

(base) [email protected]:NewProjects$ ln -s ~/Datasets/coco/* coco_kpts
(base) [email protected]:coco_kpts$ rm coco

You can delete the soft connection

2.3 function

The true annotation is shown below ：

---

YOLOPose Measured results , As shown in the figure below ：

---

Enlarged detail contrast ：

It's not hard to find out ,YOLOPose There are highlights in small-scale human detection （ As shown on the far right of the left figure , The leftmost picture on the right ）！

---

bug debugging

ImportError: cannot import name 'amp' from 'torch.cuda' (/home/wqt/anaconda3/envs/pytorch-cifar/lib/python3.7/site-packages/torch/cuda/__init__.py)

original conda In the environment torch The version is relatively low , according to install requirement.txt It is easy to have the highest version installed in , The best choice is to see the version recommended by the author , Or from issue In search of the answer .

AttributeError: Cant get attribute SPPF on module models.common

This is due to the download yolov5 Branches are relatively early , Or download the latest git, Or paste this class directly . May refer to https://blog.csdn.net/Steven_Cary/article/details/120886696

summary ：
Want to improve YOLO To improve performance , The main troubles that may be faced are ： First , Too many changes are involved ; secondly , The data format also needs to be changed to YOLO Format , If you want to run on other databases , First, we need to pass the data , A wave of worry surged into my heart

---