Masking Specific Activities in DIP-IMU Dataset

[bryanbocao]

Dear Authors,

Thanks for sharing this dataset!

Can you tell me how to index the activities in DIP-IMU?

I can only see the .pkl files organized by subjects, but how to mask or access a specific activity is unclear to me:

data/dataset_raw/DIP_IMU/s_01$ ls
01.pkl  02.pkl  03.pkl  04.pkl  05.pkl

Below are the activities from the Deep Inertial Poser paper:

It would be great to provide more clarity. Thanks!

[bryanbocao]

I understand that overview_of_published_data.txt provides the description:

Each subject contains a pickle file per motion type. For an overview of the motion types please refer to Table 6 of Appendix F. The motion types are numbered according to the category in this table (e.g. s_01/01.pkl are the upper body movements of subject 1). Note that some motions classes are split into several sub_sequences (eg. 01_a.pkl, 01_b.pkl). You can read the contents of a pickle file using this Python3 script:

import pickle as pkl
file_path = './s_01/01.pkl'
with open(file_path, 'rb') as f:
	data = pkl.load(f, encoding='latin1')
	print(data.keys())

but the fine-grained activity labels (such as Arm raises, stretches, etc.) are not clear. Sorry if I missed these information.

[kaufManu]

Hi,

As you say, the number of the file corresponds to the row number in Tab. 6 of Appendix F (ie. s_01/01.pkl are the arm raises, s_01/02.pkl the leg motion etc.). Unfortunately we do not have more fine-grained activity labels. I.e., the file s_01/01.pkl contains arm raises, then stretches, then swings, crossings etc. as described in the table, but we don't know exactly at which frames the activity changes.

You could use our visualization tool aitviewer to quickly browse through the sequences (it contains an example script how to load DIP-IMU data) and annotate the sub-actions like this.

[bryanbocao]

Got it. Thanks for your answer!

[bryanbocao]

I visualized the DIP IMU all ten subjects in https://youtu.be/QSg0yGQNrM0

The train/test split follows https://github.com/Xinyu-Yi/TransPose/blob/main/preprocess.py where s_01 to s_08 are used for training while s_09 and s_10 and used for testing. Video runs at 25fps. The order does not match exactly that in the *.pkl files, though. Hi @kaufManu, if you have time, it would be helpful to check and verify the visualization in the video for its correctness and completeness. If yes, that would be appreciated.

Vis code on Ubuntu: https://github.com/bryanbocao/Loose-Inertial-Poser/blob/main/vis_pose.py https://github.com/ZuoCX1996/Loose-Inertial-Poser/blob/main/vis_pose.py