V8heart/yolino-ttpla-benchmark

YOLinO TTPLA Benchmark (512×512)

TTPLA aerial imagery prepared for CAPSTONE / YOLinO polyline training and evaluation.

Each sample is a 512×512 PNG tile with matching NumPy (.npy) polyline labels.

Dataset structure

.
├── images/
│   ├── train/   # 905 PNG tiles
│   ├── val/     # 109 PNG tiles
│   └── test/    # 220 PNG tiles
└── labels/
    ├── train/   # 905 .npy label files
    ├── val/     # 109 .npy label files
    └── test/    # 220 .npy label files

Image and label files share the same basename (e.g. 1000_00081.png ↔ 1000_00081.npy).

Download

pip install -U huggingface_hub

# CLI (recommended)
hf download V8heart/yolino-ttpla-benchmark \
  --repo-type dataset \
  --local-dir ./YOLinO_benchmark

Python:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="V8heart/yolino-ttpla-benchmark",
    repo_type="dataset",
    local_dir="./YOLinO_benchmark",
)

Use with CAPSTONE

git clone https://github.com/V8heart/CAPSTONE.git
cd CAPSTONE
pip install -e .

export DATASET_TTPLA=/path/to/YOLinO_benchmark

bash run.sh \
  --config configs/experiments/exp77_ttpla_512512_from_exp19.yaml \
  --dataset-root "$DATASET_TTPLA"

run.sh also accepts --dataset-root directly; the training code reads the dataset root from DATASET_TTPLA or the experiment YAML field dataset_ttpla.

Related experiments

This dataset is used by the published experiment configs:

• exp77_ttpla_512512_from_exp19.yaml
• exp80_ttpla_512512_scale16.yaml
• exp81_gnn_ttpla_512512_from_exp80.yaml

Pre-trained model weights will be published separately on Hugging Face.

Citation

If you use this code or dataset packaging, please cite the original YOLinO paper and acknowledge the TTPLA dataset source used to build these tiles.