unsup_temp_embed
unsup_temp_embed copied to clipboard
Hungarian matching when #clusters is not equal to #classes
Hi Anna,
thanks a lot for uploading the code of pipeline of unsupervised learning with unknown activity names. I just want to make sure if I correctly understood your Hungarian matching in the case that the number of clusters is not equal to the number of class labels.
In the paper you mentioned “the frames of the leftover clusters are set to background”. I assume that the “background” here corresponds to the label -1
, which is only defined on the YTI dataset.
In this case, when you match the 50 clusters to the 48 ground truth classes on Breakfast, the remaining 2 clusters are simply ignored in the evaluation, as there is no background defined on Breakfast. On the YTI, you set K=9 and K’=5, which correspond to 45 clusters. In this case, there won’t be any leftover clusters but only 3 leftover ground classes. Therefore, no frames will be labeled as background during the evaluation. Only when the number of clusters on the YTI is larger than 48 will the frames of leftover clusters be assigned with label -1
.
Please correct me there is any misunderstanding. Thanks for the efforts.
Regards
Hi Wei!
Yes, you understood it totally correctly :) As I got to know that it's a really tricky question when we are talking about implementation details. I'm going to write a short 'help page' with explanations on how to use hungarian matching and what setting I was using in the work.
Thank you for the interest!
Best, Anna
On Thu, 6 Aug 2020 at 13:54, Wei Lin [email protected] wrote:
Hi, Anna
thanks a lot for uploading the code of pipeline of unsupervised learning with unknown activity names. I just want to make sure if I correctly understood your Hungarian matching in the case that the number of clusters is not equal to the number of class labels. In the paper you mentioned “the frames of the leftover clusters are set to background”. I assume that the “background” here corresponds to the label -1, which is only defined on the YTI dataset. In this case, when you match the 50 clusters to the 48 ground truth classes on Breakfast, the remaining 2 classes are simply ignored in the evaluation, as there is no background defined on Breakfast. On the YTI, you set K=9 and K’=5, which correspond to 45 clusters. In this case, there won’t be any leftover clusters but only 3 leftover ground classes. Therefore, no frames will be labeled as background during the evaluation. Only when the number of clusters on the YTI is larger than 48 will the frames of leftover clusters be assigned with label -1. Please correct me there is any misunderstanding. Thanks for the efforts.
Regards
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/Annusha/unsup_temp_embed/issues/5, or unsubscribe https://github.com/notifications/unsubscribe-auth/AB22PEF3HF5JNFAWHYYNDCLR7KKXDANCNFSM4PWQDLQA .