dinggd/Dispersion-based-Clustering

Question about the performance

c-dada opened this issue · 2 comments

c-dada commented

Hello, I set hyper-parameters 0.1 for size_penalty ,0.05 for merge_percent and 16 for batchSize in your code. But on DukeMTMC-VideoReID, the highest rank-1 is 59.3% and the heighest mAP is 47.9%. I can't reproduce the performance of your papers. Is my hyper-parameters set wrong?

LgQu commented

Hello, I set hyper-parameters 0.1 for size_penalty ,0.05 for merge_percent and 16 for batchSize in your code. But on DukeMTMC-VideoReID, the highest rank-1 is 59.3% and the heighest mAP is 47.9%. I can't reproduce the performance of your papers. Is my hyper-parameters set wrong?

Hi~ I met the same problem. However, I can get about 63% mAP on DukeMTMC-VideoReID which is higher than yours while it is still lower than the result of the paper. Besides, the mAP on MARS is 34.7% in my experiment which is so bad(43.5% in the paper). Do you try the code on MARS? How about your performance? By the way, my setting of hyper-parameters is the same as yours.

Mickeybaby commented

@ChildQuqu hello
I didn't understand this function：
def linkage_calculation(self, dist, labels, penalty):
......
linkages = linkages.T + linkages - linkages * np.eye(cluster_num)
intra = linkages.diagonal()
penalized_linkages = linkages + penalty * ((intra * np.ones_like(linkages)).T + intra).T
I don't know the details of such code.
I am a graduate student in China.Is it convenient for you to leave your mailbox or wechat?I'd like to ask you about the code details.
It doesn't matter if it's inconvenient.
Best wishes！