17,498 spatial relation triplets annotated across 3,696 images using an adversarial crowdsourcing process to minimize language and spatial biases. The dataset features 9 distinct spatial predicates and provides bounding box coordinates for both subject and object entities.
Use Cases
- Train spatial relationship classifiers using the subject and object bounding box coordinates and image pixels
- Benchmark the susceptibility of Visual Relationship Detection (VRD) models to linguistic bias using the adversarial labels
- Develop multi-modal reasoning systems that map natural language predicates to specific 2D spatial configurations
Strengths
- 17,498 labeled triplets consisting of (subject, predicate, object) relationships
- Includes bounding box coordinates for every subject and object across 3,696 unique images
- Features 9 spatial predicates including 'on', 'under', 'near', 'above', 'behind', and 'in front of'
- Incorporates adversarial examples where the spatial relationship contradicts common linguistic frequency patterns