search

waxnkw / IETrans-SGG.pytorch

This is the code of ECCV 2022 (Oral) paper "Fine-Grained Scene Graph Generation with Data Transfer".
Other
89 stars 6 forks source link

About the Augmented dataset #19

Closed liuhengyue closed 1 year ago

liuhengyue commented 1 year ago

Hi,

This is a really cool work!

I have a few question regarding the provided dataset annotation found in ModelZoo. I checked the one named I-70% + E-100%, and it contains a list 58550 entries. Does it mean these contain the modified groundtruth (GT) based on internal and external transfer? So not all GT are changed, and we should use the original annotations for those are not included in this file? For modified ones, are they the only annotations to be used for training a particular image, or do we concatenate them with the original annotations?

Also, I am still confused on 'possible_rels' even though I looked over other issues. This field is image-specific and as a list of lists of predicate labels, how do I use it? Or it is not used.

Thanks for your time!

waxnkw commented 1 year ago

Q1: Does it mean these contain the modified groundtruth (GT) based on internal and external transfer? A1: Yes.

Q2: So not all GT are changed, and we should use the original annotations for those are not included in this file? A2: No, the unchanged annotations are also included in this file. It is merged by the changed ones and the original annotation.

Q3: For modified ones, are they the only annotations to be used for training a particular image, or do we concatenate them with the original annotations? A3: it is already a merged file.

Q4: confusion about 'possible_rels'. A4: For external transfer, I add a limitation that only the relation triplet types that appeared in the training set can be added. For example, if there is no (man, drinking, horse) shown in the training set, it will be not added. The "possible_rels" is actually possible relationships between different pairs. For the final I-70% + E-100% file, it will be not used. It will only be used for generating the E100% file, where E-100% will be merged with I-70% for the final I-70% + E-100% file.

liuhengyue commented 1 year ago

Thanks for your answers!