vana77 / market-1501_attribute

27 hand-annotated attributes of Market-1501

Market-1501_Attribute

The evaluation code will be added soon.

About dataset

We annotate 27attributes for Market-1501. The original dataset contains 751 identities for training and 750 identities for testing. The attributes are annotated in the identity level, thus the file contains 28 x 751 attributes for training and 28 x 750 attributesfor test, where the label "image_index" denotes the identity. The annotations are contained in the file market_attribute.mat.

The 27 attributes are:

attribute	representation in file	label
gender	gender	male(1), female(2)
hair length	hair	short hair(1), long hair(2)
sleeve length	up	long sleeve(1), short sleeve(2)
length of lower-body clothing	down	long lower body clothing(1), short(2)
type of lower-body clothing	clothes	dress(1), pants(2)
wearing hat	hat	no(1), yes(2)
carrying backpack	backpack	no(1), yes(2)
carrying bag	bag	no(1), yes(2)
carrying handbag	handbag	no(1), yes(2)
age	age	young(1), teenager(2), adult(3), old(4)
8 color of upper-body clothing	upblack, upwhite, upred, uppurple, upyellow, upgray, upblue, upgreen	no(1), yes(2)
9 color of lower-body clothing	downblack, downwhite, downpink, downpurple, downyellow, downgray, downblue, downgreen,downbrown	no(1), yes(2)

Note that the though there are 8 and 9 attributes for upper-body clothing and lower-body clothing, only one color is labeled as yes (2) for an identity.

Sample

Evaluation

To evaluate, you need to predict the attributes for test data(i.e., 13115 x 12 matrix) and save them in advance. "gallery_market.mat" is one prediction example. Then download the code "evaluate_market_attribute.m" in this repository, change the image path and run it to evaluate.

Citation

If you use this dataset in your research, please kindly cite our work as,

@article{lin2019improving,
  title={Improving Person Re-identification by Attribute and Identity Learning},
  author={Lin, Yutian and Zheng, Liang and Zheng, Zhedong and Wu, Yu and Hu, Zhilan and Yan, Chenggang and Yang, Yi},
  journal={Pattern Recognition},
  doi = {https://doi.org/10.1016/j.patcog.2019.06.006},
  year={2019}
}

Market-1501 Dataset:

@inproceedings{zheng2015scalable,
  title={Scalable person re-identification: A benchmark},
  author={Zheng, Liang and Shen, Liyue and Tian, Lu and Wang, Shengjin and Wang, Jingdong and Tian, Qi},
  booktitle={Proceedings of the IEEE International Conference on Computer Vision},
  year={2015}
}

market-1501_attribute's Issues

how to sort the images to match the order of labels?

How to sort the images to match the images and the labels?

for example, the image:

"0665_c6s2_041293_01.jpg"

1. we convert the string "0665" to integer 665, than sort them or
2. we sort the the string "0665" in alphabet order.

In 1. "99" < "100" but In 2 "99" > "100".

I want to know which one the author used ?

License

What is the license under which the dataset is released?

Invalid values for 'up' label

I loaded gallery_market.mat and looked at the values in the third column.

The README.md says that the third column contains values for "Sleeve length" (up) and that possible values are ["long sleeve (1)", "short sleeve (2)"] but some of the data has a value of 3.

>> load gallery_market.mat
>> sum(gallery(:,3)==3)

ans =

        1234

Am I interpreting the data correctly?

Is this the right ordering of attributes?
['gender', 'hair', 'up', 'down', 'clothes', 'hat', 'backpack', 'bag', 'handbag', 'age', 'upcolor', 'downcolor']

Thanks

标签与图片的对应关系

我修改了作者的matlab代码，将图片地址与标签的对应关系存储成txt文件。 需要的可以下载：
百度网盘：
链接：https://pan.baidu.com/s/1KU8gRgQ6aaHPMXEPt6qcqQ 密码：5wkf

24 attributes in APR paper

Thanks for the work! I came across the original paper and it seems that only 24 attributes are used in the APR paper, but the dataset has 27 attributes in total. I tried to look through the paper and see if there's any specific reasons for it and did not managed to find any. May I kindly check what are the 3 attributes that are not used and if I am supposed to throw out those attributes when doing experiments?

Thank you!

Dear sir ,How can I add attribute detection to the re-id code?

Dear sir, How can I add attribute detection to the re-ID code, using the python language instead of running in matlab

is there anyway we can get a csv version?

I have tried loading the annotations in matlab and have not been able to convert them to csv, can you provide this format? thanks

Inappropriate color label

I found two images were inappropriately labeled. The color differences are calculated by CIEDE2000. RGB value standards are according to Wikipedia. I randomly picked a 3x3 area and got the mean RGB value followed by converting to L* A* B* space.

RGB_Pink = [255, 192, 203];
RGB_Purple = [128, 0, 128];
RGB_Yellow = [255, 255, 0];

1. 0430 in bounding_box_train or 227 in market_attribute.train

It was labeled as downpink = 2.
My results: de_Pink = 65.4, de_Purple = 16.3;

2. 0601 in bounding_box_test or 289 in market_attribute.test

It was labeled as upyellow = 2.
My results: de_Yellow = 85.1, de_Purple = 17.8;

Values were different according to the picked area, but the differences were significantly large.

could you share me the market_attribute.mat

how can I know the true index ?

In the mat file, the market_attribute train shape is 751x27.
How can I know the first row of it is which image ?
The bounding box of train folder have 751 identity, but i do not know how to map the label to image

Could you introduce some attribute label tools to me?

Thanks a lot, i want to label some data with attributes ,but i can't find propriate label tools to finish it.

How the data is annotated ?

Hello,
Thank you for your effort,
Could I know how you've annotate this dataset?

Do you apply yolo to crop all the people from any scene then save them as separate images and use a txt file to annotate them ?
or you've used a specific annotation tool for labeling ?

Regards

why test data(i.e., 13115 x 12 matrix)

Hi,
Based on Market-1501 dataset, we have 750 identities for test and 19,732 test images.
I understand the 12 column values, but why 13115 entries?
Thank you

implementation of the reweighthing module: how to element-wise multiplied attribute predictions score and global feature

Hi, thank you for sharing the dataset. I'm trying to implement your paper in pytorch. Got some few questions.
Let's suppose I'm using a 64 minibatch on Market-1501 dataset. And I use the 28 attributes (I personally have 30).
Using this section,

number of attribute m = 28
Equation 3). Is simply a Linear layer (ax + b) with bias term follow by a sigmoid.
for 64 minibatch, 
The attribute prediction score has shape [64, 28] i.e., R^{64xm}
The global image representation has shape [64, 2048]

In your paper, you said, you element-wise multiplied the two. How can you multiply the attribute prediction score ([64, 28]) with a feature representation of shape [64, 2048].
can you explain what I'm missing here?
Thank you.

Is that valid to use these attribute notations in MARS dataset?

Hi @vana77 , thanks for your wonderful work！

Since the MARS dataset is an extension of the Market-1501, does that mean your attribute notations also work for MARS dataset?

how to download the dataset

how to download the dataset @vana77

where can i find the train id

I have download the dataset and the attribute.mat. but i can't find the train id and test id.

The query folder of Market-1501 for what?

@vana77 ,Hi,I can not understand the query folder,gt_query folder and gr_bbox,those for what ?Thank you! and forget my poor English.

The attribute label seems not so accuracy.

As shown in following picture, the label indicates that the person doesn't have a bag. But it seems that the person has a bag.
The age seems like adult instead of teenager.
The upper-body clothing seems like pink instead of white.