Could you provide the innotation .json file for each dataset (unc, gref, gref_umd)?

[LeungWaiHo]

Dear authors:
Thanks for your work in visual grounding, I also feel interested in this field, and I have a question:
Could you provide the innotation .json file for each dataset (unc, gref, gref_umd)?

[LukeForeverYoung]

The annotations of visual grounding tasks are included in the split files. These files are released by TransVG https://drive.google.com/file/d/1fVwdDvXNbH8uuq_pHD_o5HI7yqeuz0yS/view

[tiger990111]

I found that the Flickr30k dataset have 427,193 training pairs. But it suppose to be 29,783 training pairs. Is it some mistake of split file of Flickr30k?

[LeungWaiHo]

The annotations of visual grounding tasks are included in the split files. These files are released by TransVG https://drive.google.com/file/d/1fVwdDvXNbH8uuq_pHD_o5HI7yqeuz0yS/view

Sorry, I mean the .json file for pre-training the weights of swin-Transformer.

[LukeForeverYoung]

The annotations of visual grounding tasks are included in the split files. These files are released by TransVG https://drive.google.com/file/d/1fVwdDvXNbH8uuq_pHD_o5HI7yqeuz0yS/view

Sorry, I mean the .json file for pre-training the weights of swin-Transformer.

You can generate the .json files by this script.

import torch
import json
from pathlib import Path
from icecream import ic
def get_img_set(p):
    datas=torch.load(p)
    imgs=[_[0] for _ in datas]
    return imgs

def convert_filename_4to7(name):
    return name[15:]

dataset_vt={
    'unc':['data/unc/unc_testA.pth','data/unc/unc_testB.pth','data/unc/unc_val.pth'],
    'unc+':['data/unc+/unc+_testA.pth','data/unc+/unc+_testB.pth','data/unc+/unc+_val.pth'],
    'gref':['data/gref/gref_val.pth'],
    'gref_umd':['data/gref_umd/gref_umd_test.pth','data/gref_umd/gref_umd_val.pth']

}
json_path='/mnt/ssd/Datasets/mscoco/images/annotations/instances_train2017.json'

if __name__=='__main__':
    for dataset,vt_list in dataset_vt.items():
        ic(dataset)
        img_set=[]
        for split_path in vt_list:
            img_set.extend(get_img_set(split_path))
        ic(img_set[:10])
        img_set=set(map(convert_filename_4to7,img_set))
        ic(list(img_set)[:10])
        with open(json_path,'r') as f:
            anno=json.load(f)
            for key in anno:
                ic(key)
            ic(len(anno['images']))
            ic(len(anno['annotations']))
            
            valid_img=set([img['id'] for img in anno['images'] if img['file_name'] not in img_set])

            anno['images']=[_ for _ in anno['images'] if _['id'] in valid_img]
            anno['annotations']=[_ for _ in anno['annotations'] if _['image_id'] in valid_img]

            ic(len(anno['images']))
            ic(len(anno['annotations']))
        with open(Path('coco_excluded',f'{dataset}.json'),'w')as f:
            json.dump(anno,f)