search

rxtan2 / Learning-Similarity-Conditions

46 stars 15 forks source link

The condition used in Polyvore dataset #7

Closed zuoxiang95 closed 1 year ago

zuoxiang95 commented 5 years ago

hello, @rxtan2 In your paper, I found you use 5 conditions to get the best result in the Polyvore dataset. Can you explain what conditions you used? Because I see in UT-Zappos50K, it has gender, class, closure and heel condition.

rxtan2 commented 5 years ago

Hi Zuo Xiang, the 5 condition masks do not correspond to specific masks. Instead, they are used to approximate the relevant semantic subspace for pairwise comparison. Our model is weakly-supervised and therefore, do not rely on explicitly-defined labels.

zuoxiang95 commented 5 years ago

@rxtan2 Thanks for your kind reply, I probably understand what you mean. But I still don't know how to make the triplets in the Polyvore dataset. I want to train your model in the Polyvore dataset. Can you give me any advice? Thanks again : )

zuoxiang95 commented 4 years ago

hello @rxtan2 , I recently run your model with my own dataset and the model performed well on the test dataset. When I wrote an inference, I found the model need to feed three images to get the attention weights. But when the model predicts outfit compatibility, there is not a triplet.

Charlie1994 commented 4 years ago

hello @rxtan2 , I recently run your model with my own dataset and the model performed well on the test dataset. When I wrote an inference, I found the model need to feed three images to get the attention weights. But when the model predicts outfit compatibility, there is not a triplet.

I think you can treat the averaged compatibility score of each pair of fashion items within an outfit as the final prediction score for an outfit. The implementation of their previous work, CSN [1] is quite similar to this work, in terms of compatibility score calculation.

[1] Learning type-aware embeddings for fashion compatibility, ECCV 2018

zuoxiang95 commented 4 years ago

@Charlie1994 I had seen their previous work, and I combine these two networks. It seems nice~

rxtan2 commented 4 years ago

@zuoxiang95 Hi Zuo Xiang, apologies for not checking this thread. What you can do is create attention vectors for each pair of samples. In addition, you can try out what @Charlie1994 has suggested. Btw, if it's not too much problem, do you mind emailing me if you need more clarifications. For some reason, I am not receiving email notifications for github activity.