CityPersons数据集介绍

CityPersons数据集是cityscape的一个子集，它只包含个人注释。有2975张图片用于培训，500张和1575张图片用于验证和测试。一幅图像中行人的平均数量为7人，提供了可视区域和全身标注。

CITYPERSON数据集是CITYSCAPES数据集的子集，CITYPERSON的标注文件只标注了其中HUMAN的类别。

https://github.com/CharlesShang/DetectronPYTORCH/tree/master/data/citypersons
这里的代码说了标注格式等。

CityPersons annotations
(1) data structure: 
    one image per cell
    in each cell, there are three fields: city_name; im_name; bbs (bounding box annotations)

(2) bounding box annotation format:
¡¡¡¡ one object instance per row:
¡¡¡¡ [class_label, x1,y1,w,h, instance_id, x1_vis, y1_vis, w_vis, h_vis]

(3) class label definition:
¡¡ class_label =0: ignore regions (fake humans, e.g. people on posters, reflections etc.)
    class_label =1: pedestrians
    class_label =2: riders
    class_label =3: sitting persons
    class_label =4: other persons with unusual postures
    class_label =5: group of people

(4) boxes:
¡¡¡¡visible boxes [x1_vis, y1_vis, w_vis, h_vis] are automatically generated from segmentation masks; 
      (x1,y1) is the upper left corner.
      if class_label==1 or 2
        [x1,y1,w,h] is a well-aligned bounding box to the full body ;
      else
        [x1,y1,w,h] = [x1_vis, y1_vis, w_vis, h_vis];
¡¡¡¡

这个介绍了如何将其数据格式转为voc格式
https://blog.csdn.net/weixin_38632246/article/details/98070599