Unable to reach the accuracy reported in the paper.

LZU-SIAT / AS-CAL

Augmented Skeleton Based Contrastive Action Learning with Momentum LSTM for Unsupervised Action Recognition

MIT License

31 stars 9 forks source link

Unable to reach the accuracy reported in the paper. #2

Open cjhhhh opened 2 years ago

cjhhhh commented 2 years ago

Hello,

I ran the code without modification, but unable to reach the accuracy reported in the paper. The reported result on ntu dataset with cross-subject setting is 58%, I ran the original code and only reached 49%. After I normalized the output feature vectors, the accuracy rose to 52%. Could you explain why there is such a huge gap?

Would be appreciated if you could answer this question. Looking forward. Thank you.

Mikexu007 commented 2 years ago

@cjhhhh Sorry for the late response. Different seeds will lead to different results.