Vanilla HCN accuracy

[sangwoo3]

Thanks for the neat implementation.
I've run the provided code but couldn't get the reported accuracy, 88.12%. I got 81.1%.
Have you applied any other data post-processing for NTU-D, such as normalization, other than the code (interpolation, set hip to zero)?
Have you seen a result similar to mine by any chance from your previous experiments? If so, would you give me a hint to train the network properly?

[maxstrobel]

Hi @sangwoo3,

Did you change something in the code or just clone it and run it?
I will check the issue and result deviations I find.
However, I am little busy these days, so maybe it takes a week.

Best Max

[maxstrobel]

Hi @sangwoo3 ,

I rerun the experiments from scratch and came to results very similar to my previous ones.
I updated the README and added setups for cross-view and cross-subject. Maybe you run the experiments on cross-subject?

Vanilla HCN @ 200 Epochs	Cross-View	Cross-Subject
Accuracy	89.5 %	83.8 %
Top-2 accuracy	95.65 %	91.7 %
Top-5 accuracy	98.79 %	97.2 %

I will later add updated experiments for Prototype-HCN.

If you can provide more details, maybe we can find your issue.

Best,
Max

[maxstrobel]

Since there is no response, I'll close the issue. We can reopen it if it's necessary.