您好，想请教下您，模型使用的数据集可以提供下吗？

[PKU-ICST-MIPL]

你好！ 模型使用的数据集分别是NUS-WIDE数据集和MIR-Flickr数据集，可以到数据集的官方网站下载。 NUS-WIDE：http://lms.comp.nus.edu.sg/research/NUS-WIDE.htm MIR-Flickr：http://press.liacs.nl/mirflickr/mirdownload.html

[struggling-man]

这是报的错误：FileNotFoundError: [Errno 2] No such file or directory: '../mir/feature/train_img.txt' 具体代码是在knn_mir_cross5.py文件中的这一部分（同样NUS-WIDE数据集也是缺这一部分的数据） train_img_string_list = open(feature_dir + 'train_img.txt', 'r').read().split('\n') train_txt_string_list = open(feature_dir + 'train_txt.txt', 'r').read().split('\n') train_label = open(list_dir + 'train_label.txt', 'r').read().split('\r\n')

看您的论文中提到，我的理解是在此之前利用VGG19提取了图片的4096维特征，文本方面用的是词袋模型提取的1000维特征。不知道我的理解是否有误？关于特征提取的代码以及提取后的数据，您方便共享出来吗？

[PKU-ICST-MIPL]

你好！ 图片特征使用的是vgg19模型，fc7层的特征 NUS数据集的文本特征来自于官方 MIRFlickr数据集的文本特征来自https://github.com/jiangqy/DCMH-CVPR2017

以MIRFlickr数据集为例 train_img.txt、train_txt.txt文件存放的是数据集的划分 train_img.txt的前三行样例： ++++++++++++++++ 1988.jpg 18230.jpg 16510.jpg +++++++++++++++ train_label.txt存放的是对应的标签，前三行样例： +++++++++++++++ 0 0 0 0 0 0 1 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 0 1 0 0 0 0 0 0 0 0 1 0 0 0 0 1 0 1 1 0 0 0 1 1 0 0 1 0 +++++++++++++++

[blueclowd]

Hi, the get_feature() returns a ValueError: could not convert string to float: '1988.jpg' if the extension (.jpg, .png, ...) is included in train_img.txt. Any suggestions?

[devraj89]

Hi @PKU-ICST-MIPL

Can the dataset features for the experiment be provided to us? I am unable to run the code otherwise.

Thanks Devraj

[PKU-ICST-MIPL]

Hi @PKU-ICST-MIPL

Can the dataset features for the experiment be provided to us? I am unable to run the code otherwise.

Thanks Devraj

Hi！ We use vgg19 model and adopt the output of fc7 layers as the images features The text feature of NUS datasets come from the Official dataset The text feature of MIRFlickr come from https://github.com/jiangqy/DCMH-CVPR2017

The samples of the MIRFlickr input files have been under samples directory.

[devraj89]

Hi @PKU-ICST-MIPL

Thanks for your prompt response.

I did check the https://github.com/jiangqy/DCMH-CVPR2017 link to know about the dataset. However, I have a few questions.

(1) In the paper, it is written that the text features of MIRFlickr are selected to be 1000 but from the DCMH paper, I see that 1386 dimensional BOW features are used. In the code provided also, in line https://github.com/PKU-ICST-MIPL/UGACH_AAAI2018/blob/master/UGACH-mir/train.py#L23 TEXT_DIM = 1386 Kindly please specify what was the feature dimension of the text file used?

In addition, the DCMH file has used the MIRFlickr dataset with total no of items as 20015. Whereas the paper cited [18] SePH has used MIRFlickr dataset of size 16,738. Kindly please specify what was the dataset size of MIRFlickr used for the experiments.

(2) There are only ten samples of the MIRFlickr input files provided under the samples directory. Is it possible to release all of them?

Thanking You Devraj

[devraj89]

Hi @PKU-ICST-MIPL

Thanks for clarifying the feature dimension for the MIRFLICKR text data.

I am having trouble regarding the dataset size still

(1) the SePH paper has used 16,738 as its total dataset size out of which 5000 was selected for training. (2) you have mentioned that in UGACH "We follow the split protocol of SePH, specifically, there are 5000 samples for training, 2000 samples for query and 18000 samples in the database." are used. This comes out to be a total of 25000 samples. (3) For DCMH algorithm, about 20015 samples are considered to be the total size.

kindly please clarify again the dataset that used for the experiments. I am unable to run the experiments otherwise.

Thanks Devraj

[zengdonghuo]

Hi @PKU-ICST-MIPL,

Can you release your img-txt ids for training set, database retrieval set and for query set? If so, we can easily follow your experiments and proposed new architecture less experiments with baseline.

BR Zanh

[PKU-ICST-MIPL]

Hi @PKU-ICST-MIPL,

Can you release your img-txt ids for training set, database retrieval set and for query set? If so, we can easily follow your experiments and proposed new architecture less experiments with baseline.

BR Zanh

We have release the training data of NUS-WIDE, see readme.

[PKU-ICST-MIPL]

Hi @PKU-ICST-MIPL

Thanks for clarifying the feature dimension for the MIRFLICKR text data.

I am having trouble regarding the dataset size still

(1) the SePH paper has used 16,738 as its total dataset size out of which 5000 was selected for training. (2) you have mentioned that in UGACH "We follow the split protocol of SePH, specifically, there are 5000 samples for training, 2000 samples for query and 18000 samples in the database." are used. This comes out to be a total of 25000 samples. (3) For DCMH algorithm, about 20015 samples are considered to be the total size.

kindly please clarify again the dataset that used for the experiments. I am unable to run the experiments otherwise.

Thanks Devraj

The database including the training data, the query is independent of the training data, thus the total size of MIRFlickr is 20015, which is the same as the DCMH.

[huhengtong]

Hi, Sorry I still cannot figure out the size of MIRFLICKR you used. Would you specifically give the retrieval size and the query size of MIRFLICKR? Thanks

[PKU-ICST-MIPL]

Hi, Sorry I still cannot figure out the size of MIRFLICKR you used. Would you specifically give the retrieval size and the query size of MIRFLICKR? Thanks

We adopt the MIRFLICKR features from DCMH, which have a total of 20015 samples，we select 2000 samples as the query set and 18015 as retrieval size. We select 5000 samples from retrieval set to training the supervised methods. The unsupervised methods can use all the retrieval size for feature learning, which depends on the algorithm itself.

[huhengtong]

Hi, Thanks for your prompt response. BTW, the query size is 1865 and the retrieval size is 184712 for NUS-WIDE, is that right? Thanks

[PKU-ICST-MIPL]

size

Yes, we select 1% from the nuswide for the query set.

[huhengtong]

Hi, The size of training data a of NUS-WIDE you released is 15000. So what the retrieval set size is? Is it 15000 or 184712? In addition, how many training data you used for MIRFLICKR? 15000 or 18015? Thanks

[PKU-ICST-MIPL]

Hi, The size of training data a of NUS-WIDE you released is 15000. So what the retrieval set size is? Is it 15000 or 184712? In addition, how many training data you used for MIRFLICKR? 15000 or 18015? Thanks

Our proposed method is the unsupervised method, thus the size of training data depends on the algorithm itself. We select 15000 to balance efficiency and accuracy. Only the supervised method has the constraints with 5000 samples as training set.
The size of training data for MIRFLICKR is 15000.