CrowdHuman数据集介绍

CrowdHuman数据集是旷世发布的用于行人检测的数据集，图片数据大多来自于google搜索。
CrowdHuman 数据集数据量比较大，训练集15000张，测试集5000张，验证集4370张。训练集和验证集中共有 470K 个实例，约每张图片包含23个人，同时存在各种各样的遮挡。每个人类实例都用头部边界框、人类可见区域边界框和人体全身边界框注释。
数据集可以在 http://www.crowdhuman.org/ 进行下载，其中测试数据集没有公开标注。

CrowdHuman测试
hbox：head box
vbox：可见身体box
fbox：联想扩展身体box
坐标： x y w h
cv2.rectangle：x1 y1 x2 y2
如果 x1 y1 x2 y2超出边界，不报错，但是画不出来
height,width=img.shape[:2]

import cv2
import json
 
obj_str='{"ID": "284193,faa9000f2678b5e", "gtboxes": [{"tag": "person", "hbox": [123, 129, 63, 64], "head_attr": {"ignore": 0, "occ": 1, "unsure": 0}, "fbox": [61, 123, 191, 453], "vbox": [62, 126, 154, 446], "extra": {"box_id": 0, "occ": 1}}, {"tag": "person", "hbox": [214, 97, 58, 74], "head_attr": {"ignore": 0, "occ": 1, "unsure": 0}, "fbox": [165, 95, 187, 494], "vbox": [175, 95, 140, 487], "extra": {"box_id": 1, "occ": 1}}, {"tag": "person", "hbox": [318, 109, 58, 68], "head_attr": {"ignore": 0, "occ": 1, "unsure": 0}, "fbox": [236, 104, 195, 493], "vbox": [260, 106, 170, 487], "extra": {"box_id": 2, "occ": 1}}, {"tag": "person", "hbox": [486, 119, 61, 74], "head_attr": {"ignore": 0, "occ": 0, "unsure": 0}, "fbox": [452, 110, 169, 508], "vbox": [455, 113, 141, 501], "extra": {"box_id": 3, "occ": 1}}, {"tag": "person", "hbox": [559, 105, 53, 57], "head_attr": {"ignore": 0, "occ": 0, "unsure": 0}, "fbox": [520, 95, 163, 381], "vbox": [553, 98, 70, 118], "extra": {"box_id": 4, "occ": 1}}, {"tag": "person", "hbox": [596, 40, 72, 83], "head_attr": {"ignore": 0, "occ": 0, "unsure": 0}, "fbox": [546, 39, 202, 594], "vbox": [556, 39, 171, 588], "extra": {"box_id": 5, "occ": 1}}, {"tag": "person", "hbox": [731, 139, 69, 83], "head_attr": {"ignore": 0, "occ": 0, "unsure": 0}, "fbox": [661, 132, 183, 510], "vbox": [661, 132, 183, 510], "extra": {"box_id": 6, "occ": 0}}]}'
data = json.loads(obj_str)
 
 
 
img=cv2.imread("284193,faa9000f2678b5e.jpg")
 
print(data["ID"])
for obj in data["gtboxes"]:
    # print('tag',obj['tag'])
    print('hbox',obj['hbox'])
 
    # draw_0 = cv2.rectangle(img, (obj['hbox'][0], obj['hbox'][1]), (obj['hbox'][0]+obj['hbox'][2], obj['hbox'][1]+obj['hbox'][3]), (255, 0, 0), 1)
    # print('head_attr',obj['head_attr'])
    print('fbox',obj['fbox'])
    # draw_0 = cv2.rectangle(img, (obj['fbox'][0], obj['fbox'][1]),(obj['fbox'][0] + obj['fbox'][2], obj['fbox'][1] + obj['fbox'][3]), (0, 255, 0), 1)
    print('vbox',obj['vbox'])
    draw_0 = cv2.rectangle(img, (obj['vbox'][0], obj['vbox'][1]),(obj['vbox'][0] + obj['vbox'][2], obj['vbox'][1] + obj['vbox'][3]), (0, 0,255), 1)
    # print('extra',obj['extra'])
 
cv2.imshow("'sdf",img)
cv2.waitKey()

这是如何将该数据集转换为coco数据集的操作。
https://blog.csdn.net/qq_41375609/article/details/95202218

参考链接：https://blog.csdn.net/guilutian0541/article/details/103642591
https://blog.csdn.net/jacke121/article/details/88583798