search

huhengtong / UKD_CVPR2020

The source code for the CVPR2020 paper "Creating Something from Nothing: Unsupervised Knowledge Distillation for Cross-Modal Hashing".
20 stars 12 forks source link

The construction of Simalarity #1

Closed zhangcheng-007 closed 4 years ago

zhangcheng-007 commented 4 years ago

Thank you for answering my request in QQ email. But I have some problems after reading your coding, In the setting python file of the student model, why the number of ind_img should be 8000?

huhengtong commented 4 years ago

It's a mistake, I just forget to correct it. Please follow the setting in the paper. Also, you can try other numbers to see its performance. I think the number we use may be not the optimal one.

zhangcheng-007 commented 4 years ago

Thank you for your answer. But I have another question. Is it right to say that the similarity defined in the formula can not actually calculate its real value? I used to think that we could tell if it was similar or not by the threshold s = 0.5.Thank you for your answer. But I have another question. Is it right to say that the similarity defined in the formula can not actually calculate its real value? I used to think that we could tell if it was similar or not by the threshold s = 0.5.

huhengtong commented 4 years ago

Using threshold to define similarity is another choice. We define the similarity according to the rank orders in the paper, which means the top k samples in the rank list are regarded as the similar samples with the query. I think using a soft value (the real value) of similarity may be a good try, though we use the binary value (0 or 1) in the paper.