torch==1.6.0
transformers==3.4.0
pytest==5.3.2
scikit-learn==0.22.1
scipy==1.4.1
nltk==3.4.5
We release a subset dataset (1000 samples containing). Please feel free to explore it, and the whole dataset will be released once our paper is accepted.
Steps to explore the dataset:
-
Download images from MS COCO: WebsiteMSCOCO
Please download the 2017 Train, 2017 Val, and 2017 Panoptic Train/Val.
-
The subset dataset is in the folder './benchmark', file named 'sub_ris.json'.
-
The ground truth mask of each sample can be obtained by 'get_mask.py' in the same folder. Remember to set the correct path to panoptic 2017.
The format of the annotation:
data:{
'file_name',
'sentence',
'pan_seg_file',
'id'
}
'file_name' denotes the name of the image, 'sentence' denotes the query sentence, 'pan_seg_file' denotes the name of the segmentation image, and 'id' helps to find the target area.
- Download the pre-processed data, unzip the file into '.benchmark/':
https://drive.google.com/file/d/1AOT4RMUs6nsV-rSItyTR-PGb8N-m9AhZ/view?usp=share_link
- Run the scripts below:
python train.py