search

calx-git / ST-NBNN-demo

The demo of ST-NBNN for skeleton-based action recognition
12 stars 4 forks source link

some questions about the data #1

Closed stepByStepToSky closed 6 years ago

stepByStepToSky commented 6 years ago

Hello, thanks for your kindness to do such a hard work. I get confused when I see your source code. First, the file(like /NBNN/Dataset/MHAD/a01_s01_e01_skeleton.txt) has x,y,z,and flag. I wonder how these joints are arranged(like frame1_joint1, frame1_joint2, ... frame1_join35, frame2_join1, ... frameN_join34, frameN_join35) ??? Second, I wonder why z has negative numbers ?

Thanks for help. kelly

calx-git commented 6 years ago

@stepByStepToSky

1) Yes, you are right. If there are 35 joints and 100 frames, there will be 35*100+1 lines in the skeleton file. The "1" is for recording the number of joints and frames.

2) The sign of the coordinates is actually based on the reference origin. Since each pose will be centralized by the hip joint, the sign of the coordinates does not have influence on the performance.

stepByStepToSky commented 6 years ago

@Wuie Thanks for your help. Could you please tell me how these 35 joints are arranged in your data. I means, I wonder know that each line in the file(like /NBNN/Dataset/MHAD/a01_s01_e01_skeleton.txt) represent which places in the human body.

calx-git commented 6 years ago

@stepByStepToSky

I suggest you to draw the first 35 points by scatter function in Matlab and annotate index number to them. I believe you can get the answer by yourself. 😄

stepByStepToSky commented 6 years ago

@Wuie Thanks for your replying. Actually, I follow your guidance, I find that joint13 and joint14 have the same coordinates(like joint13: 92.277540 168.469819 15.050990 1; joint14: 92.277540 168.469819 15.050990 1 in /NBNN/Dataset/MHAD/a01_s01_e01_skeleton.txt). I wonder that are there some mistakes in your data preprocessing.

Thanks for your help very much.

stepByStepToSky commented 6 years ago

@Wuie Thanks for your help sincerely. I have solved this problem. Actually, there are some noise in the dataset. Finally, thanks for your help again.

calx-git commented 6 years ago

@stepByStepToSky

Thank you for your comments. I really appreciate your meticulosity. 😁 The data of MHAD dataset is recored in BVH format, which is used for storing motion capture data in a tree structure. Actually in all the skeleton files I provide, the joint 13 and joint 14 are the same, the same situation happens for joint 20 and joint 21.

If you check the BVH files provided in MHAD dataset, you will find that for the left and right hand joints, they both have a child node (end site) recording the end position of that body part branch. The offsets from both the left and right hand to their end position node are [0, 0, 0], which means the two hands are the end of the left arm and right arm branch. So that's why you can find that both the coordinates of joint13 and joint 14 are the same. Joint 14 is actually the end position node.

Since these lines do not influence the description of human poses, we just keep them in the demo dataset. Hope my reply mollify you confusion. Thanks.