How to prepare my own dataset to train the keypoint stage?

[Mrtrl]

I wanna train delf model for product searching, however what kind of datasets should i provide for keypoint training? Is there any off-th-shelf datasets can i use for training?

[nashory]

I've got similar question via my email.

Hello ! Thanks for ur great work that I can use delf model to do a lot of things.
However, I wanna use this model for product searching (like clothes image retrieval) , so what type of dataset should I prepare for training the keypoint stage ? How should l prepare the dataset ? What should be in the dataset?

To answer, you can use any fashion datasets(deepfashion, shopping100k, DARN, ModaNet) if they provides class labels.
DeLF actually requires class-label label only for its training. (no need to have pair per instance ids since it does not use ranking loss such as triplet loss.)
From this perspective of view, you can also use instance id of image pair as a class.
So, you are free to use any datasets.

Apart from this, It's not a good idea to use DeLF on fashion domain.
Because fashion items are highly deformable which might harm the performance when performing geometric verification using RANSAC.

Furthermore, the patterns on the clothing item usually repeated, which makes the problem even harder.