Experimental Details for NTU dataset

[marco-rudolph]

Dear authors,

I read your work about "Graph Embedded pose Clustering for Anomaly Detection" with great interest.

Some questions came up for me about the experiments with the NTU dataset:

• How was the train/test split made for the non-anomalies, based on the setup or the subjects?
• What sequence length was used for the experiments?
• How were multiple subjects in the scene handled? Just as for STC, i.e., by the maximum of all subjects?
• Did you use scores on frame or clip-level?
• If you used frame-wise scores: Were the scores smoothed over the scene here as well?

Thank you in advance!

Best regards,
Marco Rudolph

[amirmk89]

Thank you for your questions, it's been a while so I'll do my best to answer these, but the paper/code remembers best :) * Train/Test split - NTU clips have a single action label per clip ("brushing hair"), this was used. A full list is in the supplementary table 8. * Entire sequence length + averaging over windows. * Multiple subjects - same as in STC. * For NTU we scored clip-level as there are no frame level annotations, if I recall correctly. No smoothing, Thank you again for your interest in our work. -Amir

…

On Mon, Apr 19, 2021 at 5:09 PM marco-rudolph ***@***.***> wrote: Dear authors, I read your work about "Graph Embedded pose Clustering for Anomaly Detection" with great interest. Some questions came up for me about the experiments with the NTU dataset: - How was the train/test split made for the non-anomalies, based on the setup or the subjects? - What sequence length was used for the experiments? - How were multiple subjects in the scene handled? Just as for STC, i.e., by the maximum of all subjects? - Did you use scores on frame or clip-level? - If you used frame-wise scores: Were the scores smoothed over the scene here as well? Thank you in advance! Best regards, Marco Rudolph — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#8>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADFQGYT5C5BYKD6CIIZNEDTTJQ2RXANCNFSM43F5D3TA> .

[Sangmann]

Dear authors,

I read your work about "Graph Embedded pose Clustering for Anomaly Detection" with great interest.

Some questions came up for me about the experiments with the NTU dataset:

• How was the train/test split made for the non-anomalies, based on the setup or the subjects?
• What sequence length was used for the experiments?
• How were multiple subjects in the scene handled? Just as for STC, i.e., by the maximum of all subjects?
• Did you use scores on frame or clip-level?
• If you used frame-wise scores: Were the scores smoothed over the scene here as well?

Thank you in advance!

Best regards,
Marco Rudolph

Hello,

I want to get the network in NTU dataset. but I couldn't apply ... Could you share your processing code?

Thanks