metadata
license: mit
ORBIT: An Object Property Reasoning Benchmark for Visual Inference Tasks
Inspired by human object categorization, this repository introduces ORBIT, a comprehensive benchmark designed to evaluate the abilities of Vision–Language Models (VLMs) to reason about abstract object properties. ORBIT spans four object property dimensions (physical, taxonomic, functional, relational), three levels of reasoning complexity (direct recognition, property inference, counterfactual reasoning), and three visual domains (photographic/real, animated, AI-generated).
Uses:
This dataset is designed to test VLMs through count-based questions, enabling evaluation of how well they reason about abstract object properties across different visual contexts.
Fields:
image_id: the image number (120 under eachimage_type).image_type: the three visual domains (photographic/real, animated, AI-generated).theme: label to indicate image theme.reasoning_complexity: the three reasoning complexity levels (direct recognition, property inference, counterfactual).question_id: the question number (three levels ofreasoning_complexity).question: the task for the specific level.answer: the ground truth/correct answer to thequestion.property_category: the four property dimensions (physical, taxonomic, functional, relational).objects: list of detected objects part of ground truth (reasoning foranswer).