基于coco数据集的人体关键点分布示意图与数据集解析

本文绘制了coco中人体姿态关键点的分布示意图，并解释了每个关键点的含义。

目录

1、数据集介绍

2、示意图

3、数据集解析

---

1、数据集介绍

        有pose标注的部分数据样式如下：

        每张图中有若干个segment标注，每个标注包含的信息如下：

{"segmentation":[[0.43,299.58,2.25,299.58,9.05,287.78,32.66,299.13,39.01,296.4,48.09,290.96,43.55,286.87,62.16,291.86,61.25,286.87,37.65,279.15,18.13,272.8,0,262.81]],
"num_keypoints":1,
"area":1037.7819,
"iscrowd":0,
"keypoints":[0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,9,277,2,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0],
"image_id":397133,
"bbox":[0,262.81,62.16,36.77],
"category_id":1,
"id":1218137}

        我们所需要的就是其中的'keypoints'部分，每三个数字为一组，代表一个关键点，三个值分别为x坐标、y坐标、标志位，其中，标志位有三个值：

• 0：未标注
• 1：标注，但被遮挡
• 2：标注，未遮挡

2、示意图

        下图中，共17个关节点（鼻子x1、眼睛x2、耳朵x2、肩部x2、肘部x2、手腕x2、髋部x2、膝关节x2、脚腕x2）：

3、数据集解析

        我们从coco2017中解析数据集并保存为YOLO格式，这种格式可以直接用YOLOv5或者YOLOv8进行训练：

"""
get person instance segmentation annotations from coco data set.
"""

import argparse
import os

import numpy as np
import tqdm
import shutil
from pycocotools.coco import COCO


def main(args):
    annotation_file = os.path.join(args.input_dir, 'annotations', 'person_keypoints_{}.json'.format(args.split))

    # init path
    subdir = args.split[:-4] + '_coco2'
    img_save_dir = os.path.join(args.output_dir, subdir, 'images')
    txt_save_dir = os.path.join(args.output_dir, subdir, 'labels')
    os.makedirs(img_save_dir, exist_ok=True)
    os.makedirs(txt_save_dir, exist_ok=True)

    coco = COCO(annotation_file)
    catIds = coco.getCatIds()
    imgIds = coco.getImgIds()
    print("catIds len:{}, imgIds len:{}".format(len(catIds), len(imgIds)))
    for imgId in tqdm.tqdm(imgIds, ncols=100):
        img = coco.loadImgs(imgId)[0]
        annIds = coco.getAnnIds(imgIds=img['id'], catIds=catIds, iscrowd=None)
        anns = coco.loadAnns(annIds)
        if len(annIds) > 0:
            img_origin_path = os.path.join(args.input_dir, args.split, img['file_name'])
            img_height, img_width = img['height'], img['width']

            lines = []
            for ann in anns:
                # if ann['iscrowd'] != 0 or ann['category_id'] != 1:
                #     continue
                bbox = np.asarray(ann['bbox'], dtype=float)  # x1y1wh
                bbox[::2] = bbox[::2] / img_width
                bbox[1::2] = bbox[1::2] / img_height
                # x1y1wh2xywh
                bbox[0] += bbox[2] / 2
                bbox[1] += bbox[3] / 2

                bbox_str = [str(b) for b in bbox]
                keypoints = np.asarray(ann['keypoints'], dtype=float)
                keypoints[::3] = keypoints[::3] / img_width
                keypoints[1::3] = keypoints[1::3] / img_height
                keypoints_str = [str(k) for k in keypoints]
                line = '{} {} {}'.format(0, ' '.join(bbox_str), ' '.join(keypoints_str))
                lines.append(line)

            if len(lines) > 0:
                txt_output_path = os.path.join(txt_save_dir, os.path.splitext(img['file_name'])[0] + '.txt')
                with open(txt_output_path, 'a') as f:
                    for line in lines:
                        f.write(line + '\n')

                img_output_path = os.path.join(img_save_dir, img['file_name'])
                shutil.copy(img_origin_path, img_output_path)


def get_args():
    parser = argparse.ArgumentParser()
    parser.add_argument("--input_dir", default="/data/public_datasets/coco2017", type=str,
                        help="input dataset directory")
    parser.add_argument("--split", default="val2017", type=str,
                        help="train2017 or val2017")
    parser.add_argument("--output_dir", default="/data/datasets/person_pose", type=str,
                        help="output dataset directory")
    return parser.parse_args()


if __name__ == '__main__':
    args = get_args()
    main(args)

---

参考：

COCO - Common Objects in Context