Datasets:

LucasSaiGao
/

src

license: cc-by-nc-4.0
task_categories:
  - image-segmentation
  - object-detection
tags:
  - topological-error-detection
  - topology
  - computer-vision
  - satellite-imagery
pretty_name: TopoAgent Dataset
size_categories:
  - 10K<n<100K

TopoAgent Dataset

This dataset contains training and validation data for the TopoAgent project, which focuses on detecting topological errors in segmentation masks.

Dataset Structure

This dataset includes the following directories:

rft_v2/ - 56.37 KB

Data Format

The dataset contains JSON files with annotations in the following format:

{
  "id": "example_id",
  "image": [
    "path/to/original.png",
    "path/to/mask.png"
  ],
  "conversations": [
    {
      "from": "human",
      "value": "<image>\n<image>\n[prompt]"
    },
    {
      "from": "gpt",
      "value": "<answer>[annotations]</answer>"
    }
  ]
}

Path Configuration

The JSON files contain absolute paths starting with /data/meilong/projects/topoagent/.

To use this dataset on your local machine:

Download the dataset
Use the replace_paths.py script to update paths:

python replace_paths.py \
    --input downloaded_file.json \
    --output updated_file.json \
    --old-prefix "/data/meilong/projects/topoagent/" \
    --new-prefix "/your/local/path/"

Error Types

The dataset includes annotations for four types of topological errors:

broken_connection: Missing connections between segments
spurious_connection: False connections in the segmentation
missing_branch: Branches that should exist but are missing
extra_branch: Branches that shouldn't exist but are present

Citation

If you use this dataset, please cite:

@dataset{topoagent_dataset,
  title={TopoAgent Dataset},
  author={Your Name},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/LucasSaiGao/src}
}

License

This dataset is released under CC-BY-NC-4.0 license.