guodong's blog

master@zhejiang university
   

h36m论文阅读

11个专业演员,女性5,男性6,BMI分布17到29,其中7个(3女+4男)用于训练验证,4个(2女+2男)用于测试。数据被分成15个场景,包含许多不对称的类型。

关节坐标和关节角度骨架表达:包括相对的3D坐标 relative 3D joint positions (R3DJP),运动学表达kinematic representation (KR)。前者坐标是相对的,相对根关节盆骨关节,根关节为坐标系的中心。后者就是各个肢体之间的相对关节角度等等信息。

此外,单目预测数据集可以通过旋转平移4摄像头操作。

附上Human3.6M的32个关键点对应的名称:

官方文件里给的结构是这样的:


names_31 = ["Pelvis", "R Hip", "R Knee", "R Ankle", "R ToeBase", "Site",
"L Hip", "L Knee", "L Ankle", "L ToeBase", "Site",
"Spine", "Spine1", "Neck", "Head", "Site",
"L Shoulder", "L Arm", "L Elbow", "L Wrist", "L HandThumb", "Site", "L_Wrist_End", "Site",
"R Shoulder", "R Arm", "R Elbow", "R Wrist", "R HandThumb", "Site", "R_Wrist_End", "Site"
]

但实际上使用时发现有几点对应错误:修改为

names_31 = ["Pelvis", "R Hip", "R Knee", "R Ankle", "R ToeBase", "Site",
"L Hip", "L Knee", "L Ankle", "L ToeBase", "Site",
"Spine", "Spine1", "Neck", "Nose", "Head",
"L Arm", "L Shoulder", "L Elbow", "L Wrist", "L HandThumb", "Site", "L_Wrist_End", "Site",
"R Arm", "R Shoulder", "R Elbow", "R Wrist", "R HandThumb", "Site", "R_Wrist_End", "Site"
]

另外,如果需要读取里面的分割图文件,需要使用h5py,可以使用:

with h5py.File(seg_path, 'r') as bs:
     segments = bs[bs['Masks'][int(seg_index)][0]]



上一篇:
下一篇:

头像

guodong

2
说点什么

avatar
1 Comment threads
1 Thread replies
2 Followers
 
Most reacted comment
Hottest comment thread
2 Comment authors
guodongAbner Recent comment authors
  Subscribe  
最新 最旧 得票最多
提醒
Abner
游客
Abner

你好,我前几天在human3.6m官网注册,想获取数据集,但是至今都没有收到回复。不知道您是否已下载human3.6m数据集,是否可以共享一份给我,谢谢。