Datasets:

Ming1010
/

NeSy-Route

license: other
license_name: openearthmap-derived-mixed-source-dataset-license
license_link: LICENSE
pretty_name: NeSy-Route
task_categories:
  - image-to-text
  - visual-question-answering
language:
  - en
tags:
  - remote-sensing
  - route-planning
  - benchmark
  - neural-symbolic
  - multimodal-large-language-models
configs:
  - config_name: task2
    data_files:
      - split: easy
        path: Task2/parquet/easy/*.parquet
      - split: medium
        path: Task2/parquet/medium/*.parquet
      - split: hard
        path: Task2/parquet/hard/*.parquet
  - config_name: task3
    data_files:
      - split: easy
        path: Task3/parquet/easy/*.parquet
      - split: medium
        path: Task3/parquet/medium/*.parquet
      - split: hard
        path: Task3/parquet/hard/*.parquet
  - config_name: task2-evaluation
    data_files:
      - split: easy
        path: Task2/evaluation/easy.parquet
      - split: medium
        path: Task2/evaluation/medium.parquet
      - split: hard
        path: Task2/evaluation/hard.parquet
  - config_name: task3-evaluation
    data_files:
      - split: easy
        path: Task3/evaluation/easy.parquet
      - split: medium
        path: Task3/evaluation/medium.parquet
      - split: hard
        path: Task3/evaluation/hard.parquet
  - config_name: task3-labels
    data_files:
      - split: train
        path: Task3/labels/labels.parquet

NeSy-Route: A Neuro-Symbolic Benchmark for Constrained Route Planning in Remote Sensing

Dataset Composition

NeSy-Route is organized into three benchmark tasks for evaluating perception, symbolic reasoning, and constrained route planning over remote-sensing imagery.

Task	Setting	Easy	Medium	Hard	Total	Content
Task 1	Few-shot	-	-	-	3,607	Semantic traversability and cost-vector prediction
Task 2	Zero-shot	7,659	3,712	1,604	12,975	Constraint-aware semantic and region reasoning
Task 3	Zero-shot	6,492	2,705	1,624	10,821	Constrained route planning with waypoint or trajectory outputs

Task 1 and Task 2 provide semantic constraint annotations for vector-based reasoning. Task 3 provides route-planning samples with endpoint annotations, optimal ground-truth trajectories, trajectory visualizations, and difficulty-level subsets.

Task 2 annotated images are packaged as Parquet shards under Task2/parquet/{easy,medium,hard}/. Each row includes the annotated image bytes and the corresponding query, region, and difficulty metadata.

Task 3 samples are packaged as Parquet shards under Task3/parquet/{easy,medium,hard}/. Each row includes the original remote-sensing image, query and endpoint metadata, the optimal ground-truth trajectory, and its visualization.

Lightweight scoring annotations are available under Task2/evaluation/ and Task3/evaluation/. Task 3 also provides 3,109 deduplicated semantic masks in a single Task3/labels/labels.parquet file. The evaluation code combines each mask with the sample-specific traversability and cost vectors to reconstruct the maps used by AR, CR, and VR.

Loading the Data

from datasets import load_dataset

task2 = load_dataset("Ming1010/NeSy-Route", "task2")
task3 = load_dataset("Ming1010/NeSy-Route", "task3")

# Lightweight ground truth used by the public scorer
task2_eval = load_dataset("Ming1010/NeSy-Route", "task2-evaluation")
task3_eval = load_dataset("Ming1010/NeSy-Route", "task3-evaluation")
task3_labels = load_dataset("Ming1010/NeSy-Route", "task3-labels", split="train")

Evaluation Code

Evaluation scripts, prompt templates, and running instructions are available in the GitHub repository:

https://github.com/MingYang1010/NeSy-Route

License

NeSy-Route follows the same licensing policy as OpenEarthMap. Label data are released under the same license as the corresponding source RGB imagery when applicable. For data whose source imagery is public domain or whose license is otherwise unspecified, the released labels follow the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC BY-NC-SA 4.0).

Please also check the LICENSE file in this repository before using or redistributing the dataset.