Wuzj/OpenRef

Towards Open-World Referring Expression Comprehension: A Benchmark with Training-free Multi-task Consistency Checker

OpenRef features three key advancements: 1) Diverse visual scenarios; 2) Variable target counts; 3) Rich vocabulary types

📂 Dataset Schema & Field Definitions

To support open-world setting in generalized REC, each entry in OpenRef follows the structured format below:

Field	Type	Description
image	string	The filename of the source image.
relation_type	string	The relationship category: single, multi, or none (for none-target cases).
proposal	list[dict]	A list of bounding boxes. Each dict contains normalized [x, y, w, h] and original_size.
positive	string	The referring expression that describes the target(s). The query expression. For single-target, multi-target cases, it refers to present target(s); for none-target cases, it describes an absent object to test model robustness.
negative	string	A distracting/hallucinated expression (primarily used in multi-target, single-target relation tasks).
bbox_num	int	Total count of target bounding boxes. 0 indicates a none-target scenario.

🔍 Field Details

• relation_type:
  • single: Exactly one target exists for the positive expression.
  • multi: Multiple instances (e.g., "yellow cars") are present; all relevant bboxes are in proposal.
  • none: The positive expression describes something not present in the image (used to test model hallucination).
• proposal:
  • Coordinates x, y, width, height are typically normalized to [0, 100].
  • original_width/height are provided to allow re-scaling to raw image pixels.
• bbox_num:
  • The bumber of bboxes in a given expression. If bbox_num is 0, proposal must be an empty list [].