Measure the diversity of image descriptions
- Python 3.6.3
- SpaCy 2.0.4
- Model:
en_core_web_sm
- Model:
- NLTK 3.2.2
- WordNet 3.0
- Matplotlib 2.1.1
- Seaborn 0.7.1
- Tabulate 0.7.7
- Numpy 1.13.1
- SpaCy 2.0.4
- Pdfcrop 1.38 (only to crop the graphs for the paper)
Instructions to either:
- Inspect our data
- Reproduce our results
- Analyze your own system
Please find all output in Data/Output/
.
First run bash get_data.sh
in the Data/COCO/Raw/
folder. This downloads the 2014 MS COCO annotation data
and unpacks the zip file. Please ensure that both JSON
files are unpacked in Data/COCO/Raw/
(and not in some subfolder). The shell script assumes you have wget
installed. If you have a Mac, this is not installed by default. You can use Homebrew to get it.
Then the following commands in order:
python annotate_coco.py
to annotate the MS COCO training and val data.python annotate_generated.py
to annotate the generated descriptions.python coco_stats.py
to generate statistics about the MS COCO data.python system_stats.py
to generate statistics about the systems.python plot_ttr_curve.py
to plot the type-token curve for MS COCO and all systems.python global_recall.py
to compute global recall.python local_recall.py
to compute local recall.python generate_main_table.py
to generate the main table.python generate_ranking_table.py
to generate the rankings.python nouns_pps.py
to generate the compound noun and PP results.python plot_compound_length.py
to generate a histogram of compound length for different systems.python plot_pp_length.py
to generate a histogram of PP depth for different systems.
If you modify any of the Python files, you can rerun the analysis using bash run_experiment.sh
.
We commented out the first two commands, because annotating all the data takes a long time.
If you are interested to reproduce our exact figures, run pdfcrop FILENAME.pdf
on the relevant files in Data/Output/
. This tool is provided with the TeXLive
LaTeX distribution.
If you don't care about other systems, you can also just run the following commands (assuming you stored your system output in descriptions.json
).
First run bash get_data.sh
in the Data/COCO/Raw/
folder. This downloads the 2014 MS COCO annotation data
and unpacks the zip file. Please ensure that both JSON
files are unpacked in Data/COCO/Raw/
(and not in some subfolder). The shell script assumes you have wget
installed. If you have a Mac, this is not installed by default. You can use Homebrew to get it.
python annotate_coco.py
python coco_stats.py
python analyze_my_system.py descriptions.json
This will first generate the basis statistics for MS COCO (the standard of comparison), and then generate all statistics for a single system. Make sure your system output is in the standard JSON format. See the Systems folder for examples.
- Bibliographic data for all systems can be found in
/Data/Systems/
. - If you use this repository, please consider citing our paper:
@InProceedings{C18-1147,
author = "van Miltenburg, Emiel
and Elliott, Desmond
and Vossen, Piek",
title = "Measuring the Diversity of Automatic Image Descriptions",
booktitle = "Proceedings of the 27th International Conference on Computational Linguistics",
year = "2018",
publisher = "Association for Computational Linguistics",
pages = "1730--1741",
location = "Santa Fe, New Mexico, USA",
url = "http://aclweb.org/anthology/C18-1147"
}