PCSR-Benchmark

PCSR-Benchmark (Perspective-Conditioned Spatial Reasoning) is a multimodal spatial reasoning benchmark covering spatial relation tasks such as CDC (Cardinal Direction Chain), designed to evaluate the directional and transitive reasoning abilities of multimodal large language models in panoramic indoor scenes.

This dataset accompanies the paper: Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images.

Dataset Summary

• Total QA pairs: 66,257 (train: 58,568 / test: 7,689)
• Scenes: 26 (from ReplicaPano)
• Images: 7,800 panoramic views
• Task types: cdc_positive, with both MCQ (multiple-choice) and Open (open-ended) formats
• Language: English

Dataset Structure

Data Splits

Split	Samples	File
train	58,568	data/data_train_58568.json
test	7,689	data/data_test_7689.json

The train/test splits are disjoint by sample id.

Data Fields

Each sample contains exactly 8 fields:

Field	Type	Description
question	string	English question
answer	string	English answer
id	string	Unique sample ID, e.g., OMM-Bench-CDC-Pos-office_3_000_00027-0
image_path	string	Relative path to the image within the repository, e.g., images/<scene_id>/<image_id>/rgb.png
scene_id	string	Scene ID
image_id	string	View ID
task_type	string	Task type, e.g., cdc_positive
options	list	Answer options for MCQ samples; empty list for Open samples

Data Instance

{
  "question": "If box is to the left of tv, and tv is to the left of wall_clock, then what is the direction of box relative to wall_clock?",
  "answer": "left",
  "id": "OMM-Bench-CDC-Pos-office_3_000_00027-0",
  "image_path": "images/office_3_000/00027/rgb.png",
  "scene_id": "office_3_000",
  "image_id": "00027",
  "task_type": "cdc_positive",
  "options": []
}

Usage

from datasets import load_dataset

# Load QA data
ds = load_dataset("Caleb-ychen/PCSR-Benchmark")
print(ds)
# DatasetDict({
#     train: Dataset({features: [...], num_rows: 58568})
#     test:  Dataset({features: [...], num_rows: 7689})
# })

sample = ds["test"][0]
print(sample["question"], "->", sample["answer"])

# Load the corresponding image
from huggingface_hub import hf_hub_download
from PIL import Image

img_file = hf_hub_download(
    repo_id="Caleb-ychen/PCSR-Benchmark",
    filename=sample["image_path"],
    repo_type="dataset",
)
img = Image.open(img_file)

Data Collection

• Scenes are sourced from ReplicaPano (a panoramic rendering of the Replica dataset), covering 26 indoor scenes with 7,800 viewpoints in total.
• QA pairs are automatically generated from spatial relations among objects in each scene, in both MCQ (multiple-choice) and Open (open-ended) formats.

License

This dataset is released under the CC BY 4.0 license. Please also check the license of the original ReplicaPano data before use.

Citation

If you use this dataset in your research, please cite:

@misc{chen2026localizationcomprehensivediagnosisperspectiveconditioned,
      title={Beyond Localization: A Comprehensive Diagnosis of Perspective-Conditioned Spatial Reasoning in MLLMs from Omnidirectional Images}, 
      author={Yuangong Chen and Wai Keung Wong and Jiaxing Li and Ioannis Patras and Xu Zheng},
      year={2026},
      eprint={2605.12413},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.12413}, 
}