liuziwei7 / fashion-detection Goto Github PK

Hello @liuziwei7,

I had looked through the paper "DeepFashion: Powering Robust Clothes Recognition and Retrieval
with Rich Annotations" but I could not find any experiments about bounding box detection. So I would like to ask something about training facts of your network.

1. It seems that you used bounding box annotations (not the landmarks ones) and trained the public fast r-cnn (I mean, a non-customized version) with the boxes. Am I on the right track?

2. Could you explain some details about training/testing data?
   (1) training data type: only in-shop(or consumer) images or both of them
   (2) the number of images: training and testing, respectively

3. Could you share your evaluation methods and results?

By the way, thanks for your sharing the models and the codes! It is really awesome and helpful!

Hi authors,

Can you please explain what was your methodology of training DARN network? DARN has n-branches for each attribute group and expects to have one label per class but in deep fashion data, many a times, multiple attributes belonging to same group get activated (multi-label per class). How are training DARN in that case?

This does not implement the FashionNet in DeepFashion: Powering Robust Clothes Recognition and Retrieval with Rich Annotations and doesn't work. Also, the parameters of layers are not well explained in the paper? can't really see how can research community reproduce the results in the paper :(

Hi Ziwei,

Thanks for your excellent idea and code along with the great contribution to the fashion field.

I meet a question after running your code and caffemodel.

It seems that the tests on **Maxi Dress** sets such as on _Painted Floral Maxi Dress_ doesn't goes well.

You can refer to the following figure：

Are you doing a further research or having an newer code for upgrading the performance? 

I‘m very interested in this project.   

Thanks!

hello @liuziwei7, I have been reading your paper Deepfashion. It is a bit confused about the loss function of attribute classification in the paper. The paper says w_pos and w_neg are two coefficients, determined by the ratio of the numbers of positive and negative samples in the dataset. My understanding is like this:

                       w_pos = num_positve / num_dataset
                       w_neg = 1 - w_pos

Is my understanding wrong? Looking forward to your reply, thank you very much!

can we detect labels of multiple people in image? How can we extend model for same?