search

98zyx / Hetero-center-loss-for-cross-modality-person-re-id

Code for paper "Hetero-center loss for cross-modality person re-identification"
MIT License
56 stars 12 forks source link

Unable to reproduce the same t-sne plots as shown in paper #3

Closed chaitrasj closed 3 years ago

chaitrasj commented 3 years ago

Hello, Thanks a lot for open sourcing your code. I have trained the SYSU-MM01 data using your code and the accuracy values are rank1 as 53.09% and mAP as 51.74% (which is almost as close to your reported values). However I am unable to get the similar t-SNE plots. I have extracted the features on trained model and reduced to 2 dimensions using sklearn.manifold tsne package. However the points are not properly clustered for same identity. Can you please share your code snippet for plotting the t-SNE? Thanks

98zyx commented 3 years ago

Hi, because I have leaved the lab, I can not give the code, but I think the reason of your problem may be the wrong arguments, because if your model are trained, at least t-sne can cluster the samples of the same identity close.

chaitrasj commented 3 years ago

Hi, The t-sne was shown for test data itself right and not for train data?

98zyx commented 3 years ago

No, I used tsne on trained data which is just for presentation.