Code for our BMVC 2020 paper.

Tested with Python 3.5 and 3.6.

python3 -r requirements.txt

1. Clone the repository

git clone https://github.com/deeplab-ai/zs-vrd-bmvc20.git
cd zs-vrd-bmvc20

2. Setup data

python3 main_prerequisites.py

3. Before training with synonyms, run notebook RelationshipSimilarities.ipynb!

The full list of tested models is included in common/models. There are two basic model categories:

• (Scene Graph) Generators: models that use standard linear classifiers.
• (Scene Graph) Projectors: models that construct linear classifiers using language.

python3 main_research.py --model=MODEL --test_dataset=TEST_DATASET

See main_research.py for other input arguments.

Example:

python3 main.py --model=visual_spat_projector

will train a visual-spatial net with a local-context-aware classifier and test it on VRD.

Other useful flags:

• compute_accuracy: compute accuracy instead of precision
• use_weighted_ce: weight cross-entropy terms
• use_merged: evaluate with merged synonym classes
• test_dataset: specify a different dataset for testing (only valid for projectors)