What is going on? as for "Instructions for fasion comming soon"

[donghaoye]

What is going on? as for "Instructions for fasion comming soon"

[AliaksandrSiarohin]

I add the instructions, tell me if something is wrong.

[donghaoye]

I think deepfashion data should be split by persion id (id_xxxxxxx), and not the persion in clothes(yyyyidxxxxxaa).

[AliaksandrSiarohin]

So you suggest to put same person in different clothes in one bucket? I think it is not correct. It is not possible to generate image in new cloth, given image in old cloth, because no information about new cloth is given to the network.

[donghaoye]

In your case, some persons will in both test and train dataset. It mixed the testing person features (face, hair, etc.. ) in the training.

[AliaksandrSiarohin]

Are you mean spiting in test and train? (I thought you mean how the pairs is created). This you right, split in test/train should be done using person_id . I think this is how it is done in new test/train split for fasion. Tell me if this is not the case.

[donghaoye]

the get_id in data/split_fasion_dataset.py should be as following: def get_id(image_path): path_names = image_path.split(r'/') path_names[-2] = pathnames[-2].replace('', '')

path_names[-1] = pathnames[-1].split('')[0]

path_names[-1] =""
return ''.join(path_names)

[AliaksandrSiarohin]

So the story is this: Initially I write script that make wrong train/test split (This one is used in arxiv paper, and it correspond to fasion-annotations-train/test). Then I contact with authors of https://arxiv.org/abs/1705.09368 . And they send me their split, that I use now (This correspond to fasion-annotations-train-new-split). And this split will be used in next version of the paper.

So now I rewrite script split_fasion_dataset.py to just move the files into corresponding folders. It is not performing actual train/test split, it just move files accordingly to already existing files train-annotations-train/test (new split). If you rewrite get_id you will have problems when creating pairs_dataset.

I hope my explanation is clear.

[donghaoye]

Thanks! I also rewrite the get_pose_name(image_path), then use the filenames of new split train/test to match the annotation file and renew the fasion-annotation-test.csv and fasion-annotation-train.csv

[AliaksandrSiarohin]

I don't understand what you mean.

[donghaoye]

I mean I will fix it by myself.

[AliaksandrSiarohin]

Ok than.