Code for the ECCV paper:
Towards Unique and Informative Captioning of Images
Zeyu Wang, Berthy Feng, Karthik Narasimhan, Olga Russakovsky
@inproceedings{wang2020spiceu,
title={Towards Unique and Informative Captioning of Images},
author={Zeyu Wang and Berthy Feng and Karthik Narasimhan and Olga Russakovsky},
booktitle={European Conference on Computer Vision (ECCV)},
year={2020},
}
- java 1.8.0+
- python 3+
- SPICE
- Parse the captions using the code from SPICE (see SPICE for the input format). Example:
java Xmx8G -jar spice-1.0.jar example_input.json -out example_parsed.json -detailed -subset
- Compute SPCIE-U score with
compute_spiceu.py. Example:
python compute_spiceu.py --parsed_input example_parsed.json --uniqueness_dict coco_train_uniqueness_dict.pkl
where the uniqueness_dict is a dictionary with key as concept and value the corresponding uniqueness score for the concept. In the paper, we calcuate the uniqueness score for a concept c using the training set with uniqueness(c) = # images not containing c / # images total. But it can be modified based on the specific application.