search

ashual / scene_generation

A PyTorch implementation of the paper: Specifying Object Attributes and Relations in Interactive Scene Generation
Apache License 2.0
187 stars 30 forks source link

The number of train set in code is not identical with paper #6

Closed ghost closed 4 years ago

ghost commented 4 years ago

Hi, I have found in your paper you said the train set is approximately 25,000 but in code, there are 72,623.

72,623 comes from len(self.image_ids) in coco.py. 25,000 comes from paper.

Does anyone know anything about it?

ghost commented 4 years ago

Ok, I get it. The reason behind this is that stuff_annotations_trainval2017.zip has been changed. The original version has a 40K train set and a 5K val set which is a subset of all coco images. But after 06/17/2018, the stuff_coco has been completed, that means it has all annotations for all image of coco trainval2017. These refs from here If you want to find the original version, please download stuff_annotations_trainval2017.zip first and then unzip it. Finally, use the annotations/deprecated-challenge2017/train/val-ids.txt to find the original version.