EQA Dataset

EQA v1.0

Download (6.5M)
9,000 questions from 774 environments

Readme

v1.0 contains location, color, and place preposition questions
Includes bounding box annotations for target object and room
The train, val, test splits have no overlap in environments
Numbers (in papers, etc.) should be reported on v1.0 test

Format

{
  'splits': {
    'train': [
      '04ae3acd3052cd841158c793d2aa7ab2',
      'b73157d73cfa475e4d6bf486332d7e3a',
      ...
    ],
    'val': [...],
    'test': [...],
  },
  'questions': {
    '04ae3acd3052cd841158c793d2aa7ab2': [
      {
        'id': 296682,
        'question': 'what room is the playstation located in?',
        'answer': 'living room',
        'type': 'location',
        'bbox': [
          {
            'box': {
              'max': [51.46499584966768, 0.6355870163111009, 30.724973313243318],
              'min': [51.17500185614954, 0.5217500136674362, 30.435025319724154]
            },
            'name': 'playstation',
            'target': True,
            'type': 'object'
          },
          {
            'box': {
              'max': [52.47473515522324, 2.739999938756229, 32.65480191177438],
              'min': [46.33526263620115, 0, 27.4647918665828]
            },
            'name': ['living room'],
            'target': False,
            'type': 'room'
          }
        ]
      },
      ...
    ],
    ...
  }
}

EQA-Humans v1.0

Navigation trajectories and answers collected from humans on AMT for 1,162 questions across 70 environments are available here.