TDIUC stands for Task Directed Image Understanding Challenge, which is a visual question answering dataset. The dataset includes 1.6 million questions and 170,000 images from MS COCO and Visual Genome datasets; image-question pairs are divided into 12 categories and 4 additional evaluation matrices to evaluate model robustness and reasoning ability.

TDIUC statistics:

• 12 different question types (grouped by basic tasks, including some "absurd" questions)
• 167,437 images (from MS-COCO and Visual Genome)
• 1.65 million question-answer pairs (from 3 sources)
• 4 new evaluation metrics (designed to compensate for bias)
• 6 new experimental settings (answering key questions about VQA algorithms)