README.md

metadata

license: cc-by-nc-sa-4.0

PIO Dataset Card

Resources

• [Github] https://github.com/PIO-benchmark/pio-benchmark

Dataset Details

This dataset contains the dataset for NeurIPS D&B submission: Point-It-Out: Benchmarking Embodied Reasoning for Vision Language Models in Precise Visual Grounding. It includes all the data for S1, S2 and S3 defined in the paper:

• 260+ source images
• 550 + questions
• Human annotated masks (for S1 ans S2)
• class information and subclass information

Dataset Structure

• images_s1s2 folder
  • Contains the raw images for s1 and s2
• images_s3 folder
  • Contains the raw images for s3
• `s1s2.json'
  • List of data points include dict_keys(['polygon', 'image_path', 'height', 'width', 'lang', 'dataset_name', 's1_or_s2', 'subclasses'])
• 's3.json'
  • A dict, the key is source image, and value is List[str] including multiple questions regarding this image

Citation

None, paper in submission