Semantic Propositional Image Caption Evaluation (SPICE)
Evaluation code for machine-generated image captions.
Requirements
- java 1.8.0+
Dependencies
- Stanford CoreNLP 3.6.0
- Stanford Scene Graph Parser
- Meteor 1.5 (for synset matching)
Usage
To run SPICE, call the following (from the target directory):
java -Xmx8G -jar spice-*.jar
Running SPICE with no arguments prints the following help message:
SPICE version 1
Usage: java -Xmx8G -jar spice-*.jar <input.json> [options]
Options:
-out <outfile> Output json scores and tuples data to <outfile>
-cache <dir> Set directory for caching reference caption parses
-threads <num> Defaults to the number of processors
-detailed Include propositions for each caption in json output.
-noSynsets Disable METEOR-based synonym matching
-subset Report results in <outfile> for various semantic tuple subsets
-silent Disable stdout results
See README file for additional information and input format details
The input.json file should contain of an array of json objects, each representing a single caption and containing image_id, test and refs fields. See example_input.json
It is recommended to provide a path to an empty directory in the -cache argument as it makes repeated evaluations much faster.
Build
To build SPICE and its dependencies from source, and run tests, use Maven with the following command: mvn clean verify. The jar file spice-*.jar will be created in the target directory, with required dependencies in target/src.
Building SPICE from source is NOT required as precompiled jar files are available on the project page.
A note on the magnitude of SPICE scores
On MS COCO, with 5 reference captions scores are typically in the range 0.15 - 0.20. With 40 reference captions, scores are typically in the range 0.03 - 0.07. This is the expected result due to the impact of the recall component of the metric. To make the scores more readable, on the MS COCO leaderboard, C40 SPICE scores are multiplied by 10.
Policy gradient optimization of SPICE
We read with interest a paper that directly optimized SPICE (and other metrics) using policy gradients. The results indicated that optimizing SPICE and CIDEr (SPIDEr) produced the best captions, but that optimizing SPICE on its own leads to ungrammatical results. This is because SPICE ignores, and does not penalize repeated scene graph tuples. However, it would be straightforward to adjust the metric to penalize repetition. Contact us for details.
References
If you report SPICE scores, please cite the SPICE paper:
Developers
- Peter Anderson (Australian National University) ([email protected])
Acknowledgements
- This work is based on the SceneGraphParser developed by Sebastian Schuster (Stanford).
- We re-use the Wordnet synset matching code from Meteor 1.5 to identify synonyms.
License
- GNU AGPL v3