This repository is part of the codes for our paper
1. Download raw dataset files
Due to the maximum file size limitation of GitHub, not all of the dataset files needed are included.
- Download captions_train2017.json, captions_val2017.json, and instances_train2017.json on http://cocodataset.org/#download. Put the first two files at annotations/MSCOCO (you need to create a folder named 'MSCOCO'), and put the last file at annotations/Objects.
- Download train_places205.csv and val_places205.csv on http://data.csail.mit.edu/places/places205/trainvalsplit_places205.tar.gz. Put them at annotations/Scenes.
2. Create results folders
GitHub ignores empty folders when uploading files, so please create a folder named 'results', and create three folders in this folder named 'description_analysis', 'img_num_analysis' and 'VQA'. Alternatively, you can modify the locations of result files in the codes.
3. Run the codes
- Run prepare_dataset_comparison.py and prepare_dataset_comparison_2.py (must in Python 2), and then see compare_datasets.ipynb to reproduce the result of Figure 2 in our paper.
- Run description_analysis_mscoco.py to reproduce some of the results of Table 1 in our paper. The numbers may slightly differ from printed in our paper due to NLTK's features.
- Run VQA_analysis.py to reproduce the result of Table 3 in our paper.