You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

TunnelWeave-Segmentation

8,889 original training images, including 8,809 real images and 80 existing synthetic supplementary images. Original mask parts and validity masks are preserved byte-for-byte. For a class, OR its mask_parts and intersect with valid_region; missing parts mean zero target only within covered classes. A null valid_region means the whole image for full_image_coverage_classes. Other classes and pixels remain unknown. presence uses 1=present, 0=confirmed absent in a fully covered class, -1=unknown. ROI-zero is not global absence.

Organization

Native images and annotations are embedded in Parquet shards. samples.jsonl identifies each row. index--*.jsonl indexes all matching examples. Multilabel segmentation images occur in every relevant category index while the image is stored only once (the first positive class is the storage-file category prefix, not an exclusive label). scoped_empty is not an all-class negative label; normal is reserved for explicit full thirteen-class normal coverage.

from datasets import load_dataset
ds = load_dataset("augustusper/TunnelWeave-Segmentation", "all", split="train", token=True)

Source and usage

Source datasets: TACK (https://arxiv.org/abs/2512.14477), dacl10k (https://openaccess.thecvf.com/content/WACV2024/html/Flotzinger_dacl10k_Benchmark_for_Semantic_Bridge_Damage_Segmentation_WACV_2024_paper.html), RFDD (https://doi.org/10.1038/s41597-026-07851-7), the CTCD release used by the project, Mendeley Water leakage (https://doi.org/10.17632/xz2nykszbs.1), and author-collected tunnel imagery. See per-record source_dataset. Original source terms continue to apply. This private research archive does not establish a blanket redistribution license; public release requires source and author permissions. No test, validation, or cross-tunnel diagnostic split is included.

Source and synthetic provenance are retained per record. Images are not resized or regenerated. No private file-system paths, credentials, run logs, or model weights are included. export_summary.json and SHA256SUMS record the package contents.

Downloads last month
191

Paper for augustusper/TunnelWeave-Segmentation