xiaoxiaoguo / fashion-iq

fashion-iq

About this repository

Fashion IQ is a dataset we contribute to the research community to facilitate research on natural language based interactive image retrieval. We released Fashion IQ dataset at ICCV 2019 workshop on Linguistics Meets Image and Video Retrieval.

The images can be downloaded from here.

The image attribute features can be downloaded from here.

Starter code for Fashion IQ challenge

To get started with the framework, install the following dependencies:

• Python 3.6
• PyTorch 0.4

Citations

If you find Fashion IQ useful, please cite the following paper:

@article{guo2019fashion,
  title={The Fashion IQ Dataset: Retrieving Images by Combining Side Information and Relative Natural Language Feedback},
  author={Wu, Hui and Gao, Yupeng and Guo, Xiaoxiao and Al-Halah, Ziad  and Rennie, Steven and Grauman, Kristen and Feris, Rogerio},
  journal={CVPR},
  year={2021}
}

License

Community Data License Agreement (CDLA) License

fashion-iq's Issues

How to divide the data set into training set and test set

> The previous link for attributes was broken for some reason. So updated the link for attributes.

    > The previous link for attributes was broken for some reason. So updated the link for attributes.

It broke again, please help, thanks!

1.The URLs in "first link" have many broken links, could you please upload the dataset'images in a more stable way? Thanks a lot!
I download through the URLs, there are hundreds of picture show 0byte, which means I failed to download the picture through URLs.

2.The files in second link are disappeared.

3.The third link cannot been accessible.

The first question is very urgent for me, and also images in a dataset is undoubtedly very significantIt, so please upload the dataset'images in a more stable way. Thanks and thanks a lot!!

I'll appreciate you if you do that.
May the Force be with you. Thanks again!

Originally posted by @kennyorn1 in #17 (comment)

Hyper-parameter settings

Hi, Xiaoxiao:

 Thanks for you work and repos.  I try to rerun your model reported in the CVPR <Fashion IQ: A New Dataset Towards

Retrieving Images by Natural Language Feedback>.

 It's an interesting and powerful model. But it cannot achieve the same good performance when I run this repo. Specifically, the  performance of relative captioning implemented based on this repo is  far worse than the performance reported in the paper. I think the hyper-parameter may be the key.  So would you like to share the detailed hyper-parameters setting of the experimental result reported in your paper? 

 Thanks very much.

How can I download the dataset?

The links given in the README is not working.

New dataset download link

Due to the link in the read.me is taken down, I reshare it at https://disk.yandex.com/d/Z2E54WCwvrQA3A

if the link is not availble, plz contact me at [email protected].

resize images issues

i have download the image using image url。subsequent， should i seize the image before the torchvision transforms？ cause some image size is less than the crop size, and causing transfom
error。or should i change the crop size more lesser？

Collating captions produces zeros

These lines apparently copy zeroes from the newly created tensor instead of actual token ids:

fashion-iq/start_kit/data_loader.py

Lines 89 to 92 in 529a8b9

	captions = torch.zeros(len(captions), max(lengths)).long()
	for i, cap in enumerate(captions):
	end = lengths[i]
	captions[i, :end] = cap[:end]

Pretrained user models for relative captioning

Can you share a checkpoint for the User Model?

Two links here cannot be accessible!!!, which terribly hindered my research work!!!

Two links here cannot be accessible!!!
Please, please, plea........se!! repair the problem as soon as possible. Because of this problem, my research work was dramatically, terribly hindered(keep crying wuwuwu...)
So please, don't let us down

I'll appreciate it if this problem is handled. Thanks and thanks a lot!!!

The image attribute features cannot be downloaded

The address given in README is not available:
https://ibm.box.com/s/imyukakmnrkk2zuitju2m8akln3ayoct

How to solve the problem of broken link in fashioniq

Hello,
some url links are broken in fashioniq. How can I get the right pictures?

Broken URLs

I found some of the URLs provided in asin2url.{dress, shirt, toptee}.txt are broken, but these broken images still remain in the image_split file, which causes errors.

Could you please provide the images set that matches the image_split file?

about attribute predict

i don't see the image attribute dataset，can the attribute data find in the deepfashon data？please let me know。

attribute features

The link of attribute features image has expire or down

File readers

Dear Xiao,

Thanks for sharing the file readers, and the models.

Large amount of image urls are broken

Hi，
However, I try to download the dress data of FasionIQ dataset but found that about 905 image URLs are missing. I finally get 18182 dress image data. I hope to know whether you get a complete dataset and how many images are included in the dataset.
Can you send the datasets to google drive?
Thanks.

Some logo pictures in dataset

I notice there are some pictures that are not clothing, are they normal?
For example, ID B00APYBNR0 is not a dress, but in cap.dress.train.json it says:

    {
        "target": "B00APYBNR0",
        "candidate": "B00A890DQY",
        "captions": [
            "is a short sleeved floral dress",
            "is a company name and not a dress"
        ]
    },

It seems the first caption is wrong?