A collection of multi-turn guessing games utilizing VisualGenome images and scene graphs for attribute grounding tasks. The data serves as a multi-task framework to evaluate the quality of neural representations through object identification and visual dialogue.
Use Cases
- Evaluate attribute grounding by identifying target objects within a VisualGenome reference scene
- Assess the quality of neural representations using the provided multi-task evaluation framework
- Train models to perform object identification based on descriptive attributes in a game-based setting
Strengths
- Uses images and reference scenes sourced from the VisualGenome dataset
- Focuses on attribute grounding within a multi-task evaluation framework
- Structured as a series of goal-oriented games for neural representation testing