Dataset Card for TOPVIEWRS

The TOPVIEWRS (Top-View Reasoning in Space) benchmark is a multimodal benchmark intended to evaluate the spatial reasoning ability of current Vision-Language Models. It consists of 11,384 multiple-choice questions with either realistic or semantic top-view map as visual input, across 4 perception and reasoning tasks with different levels of complexity. For details, please refer to the project page and the paper.

Dataset Details

Dataset Features

• Multi-Scale Top-View Maps: Multi-scale top-view maps of single rooms and full houses add divergence in the granularity of the entities (objects or rooms) in spatial reasoning.
• Realistic Environmental Scenarios with Rich Object Sets: Real-world environments from indoor scenes, with 80 objects per scene on average.
• Structured Question Framework: Four tasks including 9 sub-tasks in total, allowing for a fine-grained evaluation and analysis of models’ capabilities from various perspectives and levels of granularity.

Dataset Statistics

The TOPVIEWRS evaluation dataset comprises a total of 11,384 multiple-choice questions after human verification, with 5,539 questions associated with realistic top-view maps, and 5,845 with semantic top-view maps. The choices are uniformly distributed over choices A(25.5%), B (24.6%), C (24.5%) and D (25.4%).

The maps are collected from Matterport3D dataset, which includes 90 building-scale scenes with instance-level semantic and room-level region annotations in 3D meshes. We filter these to exclude multi-floor and low-quality scenes, selecting 7 scenes with an average of 80 objects and 12 rooms each.

Note: We only release part of the benchmark (2 different scenarios covering all the tasks of the benchmark) in this dataset card to avoid data contamination. For full access to the benchmark, please get in touch with Chengzu Li via email: [email protected]

Uses

data = load_datasets(
        "chengzu/topviewrs",
        trust_remote_code=True,
        map_type=MAP_TYPE,
        task_split=TASK_SPLIT,
        image_save_dir=IMAGE_SAVE_DIR
)

To use the dataset, you have to specify several arguments when calling load_datasets:

• map_type: should be one of ['realistic', 'semantic']
• task_split: should be one of ['top_view_recognition', 'top_view_localization', 'static_spatial_reasoning', 'dynamic_spatial_reasoning']
• image_save_dir: specify the directory where you would like the images to be saved

Data Instances

For example an instance from the top_view_recognition task is:

{
  'index': 0,
  'scene_id': '17DRP5sb8fy',
  'question': 'Which of the following objects are in the room?',
  'choices': ['shelving', 'bed', 'toilet', 'seating'],
  'labels': ['bed'],
  'choice_type': '<OBJECT>',
  'map_path': '<IMAGE_SAVE_DIR>/data/mp3d/17DRP5sb8fy/semantic/17DRP5sb8fy_0_0.png',
  'question_ability': 'object_recognition'
}

Data Fields

Every example has the following fields

• idx: an int feature
• scene_id: a string feature, unique id for the scene from Matterport3D
• question: a string feature
• choices: a sequence of string feature, choices for multiple-choice question
• labels: a sequence of string feature, answer for multiple-choice question. The label's position in the choices can be used to determine whether it is A, B, C, or D.
• choice_type: a string feature
• map_path: a string feature, the path of the input image
• question_ability: a string feature, sub-tasks for fine-grained evaluation and analysis

For dynamic_spatial_reasoning task, there would be one more data field:

• reference_path: a sequence of list[int] feature, the coordinate sequence of the navigation path on the top-view map.

Citation

@misc{li2024topviewrs,
      title={TopViewRS: Vision-Language Models as Top-View Spatial Reasoners},
      author={Chengzu Li and Caiqi Zhang and Han Zhou and Nigel Collier and Anna Korhonen and Ivan Vulić},
      year={2024},
      eprint={2406.02537},
      archivePrefix={arXiv},
      primaryClass={cs.CL}
}