knn_pose_classification_create_csv

Using MediaPipe Pose Classification soultion to get 33 landmarks of body which create and validate a training set for the k-NN classifier, and export result to a CSV. The purposeis is use it in the ML Kit sample app.

There are two versions for usage. Version 1 is the use command to run, version 2 is running with user interface to do.

Folder structure:

├───log
├───resource
│ ├───extract_images
│ └───src_video

Video limit:
1.Pose type prefer counting style.
2.Video length is around 30 seconds.
3. Input video view only have one person.

Usage version 1

• Step 1:
  Select the a pose video and put it in ./resource/src_video location for prepossing. And extract the video become to images by calling the extract_images() function in extract_images.py. And run the code:

  	python extract_images.py   

• Step 2:
  In location ./resource/extract_images will find a [pose_class] folder. Select what you need to classify a pose label like up-and-down which manual pick images into two folders that named pose_class_up, pose_class_down. And delete the original [pose_class] folder.

  E.g., [pose_class] is a folder name 'jumping_jacks'. Manual pick up images into those folders from jumping_jacks folder by   
  creating two folders named 'jumping_jacks_up' and 'jumping_jacks_down' in './resource/extract_images/' location.   
  
  Finally, rember to delete 'jumping_jacks' folder. The structure will look like:   
  
  ├───resource   
  │   ├───extract_images   
  │   │   ├───jumping_jacks_down   
  │   │   └───jumping_jacks_up     
  

• Step 3:
  Modify output path location variables in csv_create.py file. The variable, export_csv is a pose classify CSV file with KNN method result. When modify variables done, we can run the code:

  	python csv_create.py  

Usage version 2

• Run the code:
  We will see the all functions for version 1 from extract images to export CSV files with GUI.

  	python csv_create_GUI.py  

• Step 1:
  Input the words create a pose, and click the button 'Extract pose'. That will help you to select the source video and extract video become to images in a renamed (your input words) folder. If extract progress is made, the status will hint 'Extract done!'.

• Step 2-1:
  Click the 'Path select' button, and select the extract images folder. The target path will hint the complete path ./resource/extract_images.

• Step 2-2:
  This step creates two status label on a pose, e.g., up-and-down. To 'Submit' button to create two states in 'pose 01' and 'pose 02', and notice the pose state words must add '_' string in middle between poses and status. The pose word just follow the step 1 you created.

• Step 2-3:
  If made the step 2-1 to step 2-2 done, we can click 'Classify' button go to the new window 'pose classify'. E.g., In step 1 create jumping jack pose, and step 2-2 input the jumping jack_up and jumping jack_up to submit folders. After press 'Classify' button, we will go to new window. Click Check Status button to refresh images display. And would be see the two button jumping jack_up and jumping jack_up.
  You can see the image to decide pose status (click up or down button) and then remember click the 'Check Status' button to refresh the image panel. Press 'Check Status' button will close the window and go back the export window when images status all you classified. The new window will see as the follow:

• Step 3-1 and Step 3-2:
  Click 'Imgs log' and 'CSVs log' buttons to set path log. This steps are preparing for step 3-3 export result CSV file to debug.

• Step 3-3:
  Key the output words to name an export file in 'Set file'. Notice the file format is '.csv'. If all progresses complete made, the status will hint extract done detail information.

The usage version 2 demo as the follow:

File description

extract_images.py

Function describe as the below:

camera_info(): To test basic camera view.

imgs_to_video(): Collect images which in a specific folder and convert to a video.

extract_images(): Input a video(or camera) to save extract images to [pose_class] folder.

csv_create.py

Just modify output path location variables:

bootstrap_images_in_folder: Specific folder (e.g., extract_images) need have at least a pose class or many different pose class folder.

bootstrap_images_out_folder: Output folders for bootstrapped images log.

bootstrap_csvs_out_folder: Output folders for bootstrapped CSVs log.

export_csv: The finally output pose label csv file we want.

Install

Conda virtual env

conda create --name [env_name]  python=3.8
conda activate [env_name]
pip install -r requirements.txt