Convolutional neural networks are widely used in image classification. But perform badly when it is an abstract rule like identity or symmetry. In this dataset we conducted a study with humans on three different datasets based on abstract rules. In addition to the study we used an eye tracker to gather data of participants' eye movements.

• 13 participants classified 12 selected tasks in the same order
• 12 tasks consisting of generated and randomly selected images:
  • 2 PSVRT tasks: SR and SD
  • 4 SVRT tasks: Problems 1, 19, 20 and 21
  • 6 Checkerboard tasks: 2 fixed camera position, 2 random board placements and 2 camera rotated on a sphere
• Eye movements were tracked with a Tobii X2-60 eye tracking device, satisfying the recommended distances
  • Gaze data was saved as Gazedata objects from Tobii
• Publicly available to Download (~270MB).

Svrt 1, 19, 20 and 21

PSVRT with same/different bitpatterns, each once horizontally and once vertically aligned

Fixed camera position and random placements of checkerboards

Camera rotated on a sphere

To be published

This research was possible due to the Management Center Innsbruck providing the eye tracking device.

rick.spiegl@student.uibk.ac.at