mabatin/pose_detection

ML intern task for Headless Technologies Ltd.

Introduction

Keypoint extraction is done using mmpose pose detection library.

For dataset, download from drive_download and extract like this:

pose_detection
|── data
    |── images
        |-- Image_1.jpg
        |-- Image_4.jpg
        |-- ...
    |── labels.csv

Run script imabalance.py for creating unbalanced_keypoints.csv.

unbalanced_keypoints.csv file format is:

image_file_name, <x_1, y_1>, <x_2, y_2>, ..., <x_n, y_n>, label

where each <x_n, y_n> corresponds to the coordinate of a keypoint. Each image has a total of 17 keypoints. body/2d_kpt_sview_rgb_img/topdown_heatmap

Run script balance.py for creating balanced_keypoints.csv.

Balancing is done by using RandomOverSampler from imblearn to upsample minority labels.

Resultant .csv files are saved in

pose_detection
|── results
    |-- balanced_keypoints.csv
    |-- unbalanced_keypoints.csv

Run script misc.py for counting labels for each csv file