KNN with k=1 or 9?

[ChengYiBin]

Is this the setting for NTU RGB+D 60 Dataset?

neigh = KNeighborsClassifier(n_neighbors=9, metric='cosine')

code in this line

[DragonLiu1995]

For Cross-View, we used n_neighbours=1, but for Cross-Subject we actually tried a set of parameters(1, 3, 5, 7, 9). This code snippet was cleaned up from the jupyter notebook we did for Cross-Subject experiment.

[ChengYiBin]

So for Cross-Subject, k=9 is the best hyper-parameter for action recognition, and is selected for all Cross-Subject evaluation, right ?

[DragonLiu1995]

Yes.

[ChengYiBin]

sorry to bother again. About the experimental results in this paper, those unsupervised method compared with this work, I wonder whether you re-implement them according to original papers (i.e. fine-tune a final FC layer), or also by KNN classification?
thanks!

[DragonLiu1995]

Except for LongT GAN which is implemented by us, the other results come from their papers.

[ChengYiBin]

thank u for reply : )