Hugging Face's logo

Datasets:
openconstruction
/
ConSynth-X

Tasks:
Object Detection
Image-to-Image
Visual Question Answering
Modalities:
Tabular
Text
Formats:
parquet
Languages:
English
Size:
10K - 100K
Tags:
construction
synthetic-data
weather-augmentation
domain-shift
robustness
vlm-benchmark
DOI:
Libraries:
Datasets
pandas
Polars
License:
Dataset card Data Studio Files
xet
 Community
1

ConSynth-X — Access Request

ConSynth-X contains synthetic augmentations derived from Construction-Site (CC BY 4.0) and SODA (research-only) datasets. Access is granted on a per-request basis after review by the OpenConstruction team. Approval is typically within 5 business days.

By requesting access you agree to: (1) use the data for non-commercial research only; (2) cite the ConSynth-X paper in any derived work; (3) for any SODA-derived rows (soda_voc, soda_ktsh subsets), separately obtain permission from the original SODA dataset team before redistribution or publication of derived artifacts.

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

ConSynth-X (v2-strict on main)

ConSynth-X is a construction computer-vision robustness benchmark: paired synthetic augmentations of real construction site images under weather, lighting, scale, and domain-shift conditions, for evaluating object detectors and vision-language models (VLMs).

This repository hosts the v2-strict release (DINO ≥ 0.85, 34,199 synthetic rows across 3,109 source scenes × 11 conditions).

Preview grid (5 source images × 11 synthetic conditions) 5 source images shown across the 11 synthetic conditions — bounding boxes drawn for the cs10k subset.

⚠️ original split — source image placeholder (not a synthetic condition)

The original split is not one of the 11 distributed synthetic conditions. It is a GT-only placeholder: rows contain full annotations (bounding boxes, captions, all metadata) but the image column is null because upstream dataset licenses prohibit redistribution of source images. The 11 synthetic conditions ship complete image bytes and are unaffected.

To reconstruct the source images locally, see Reconstructing source images (Bring-Your-Own-Source) below. A one-command rebuild script and per-image sha256 manifest are included in this repository (rebuild_consynthx.py, original_manifest.parquet).

v2-strict summary

Subset Source scenes Synthetic rows (× 11 conditions)
cs10k 1,586 17,446
soda_voc 608 6,688
soda_ktsh 915 10,065
Total 3,109 34,199

Every source scene is fully paired across all 11 synthetic conditions, enabling rigorous cross-condition (paired) robustness analysis.

Gating rule. A source image is retained iff all 9 GATE conditions (fog_{light,medium,heavy}, rain_light, snow_{light,heavy}, night, small) achieve DINO cosine similarity ≥ 0.85 against the source. Three conditions are autopassed (rain_heavy, rain_night, snow_night) because their physics-overlay pipelines were already quality-controlled upstream.

DINO model. facebook/dinov3-vitl16-pretrain-lvd1689m — CLS token, L2-normalized cosine.

Synthetic conditions (11)

Group Conditions
fog fog_light, fog_medium, fog_heavy — Depth Anything V2 + Koschmieder physics
rain rain_light, rain_heavy — IP2P + physics overlay
snow snow_light, snow_heavy — IP2P + physics overlay
night night — CycleGAN-Turbo day→night
night × weather rain_night, snow_night — Order B: orig → IP2P → day2night → physics
scale small — FLUX.1-Fill-dev outpainting

Source datasets

Subset Source License Notes
cs10k Construction-Site (LouisChen15) CC BY 4.0 Bounding boxes (objects field, VOC schema)
soda_voc SODA-A (VOC view) research-only Bounding boxes; derived users must obtain permission from the SODA team
soda_ktsh SODA (caption/KTSH view) research-only Captions; derived users must obtain permission from the SODA team

Schema (per parquet row)

  • image — JPEG image bytes. Null for the original split (placeholder; reconstruct with rebuild_consynthx.py). Full bytes for all 11 synthetic splits.
  • image_id, source_id, source_dataset
  • condition, condition_labels (multi-label)
  • objects (cs10k / soda_voc — VOC bbox list, normalized [0,1]) or captions (soda_ktsh)
  • rule_*_violation (cs10k — safety rule flags)
  • pipeline (generator metadata), quality_scores (DINO, etc.)

Pair across conditions by source_id.

Loading 

from datasets import load_dataset

# v2-strict (default = main) — synthetic split, full image bytes
ds = load_dataset("openconstruction/ConSynth-X", "cs10k", split="fog_heavy")

# All conditions for a subset
ds = load_dataset("openconstruction/ConSynth-X", "soda_voc")

Note: loading split="original" returns rows with image=None. Reconstruct first (see below), then load from the local rebuilt tree.

Reconstructing source images (Bring-Your-Own-Source)

The original split carries full GT (bounding boxes, captions) but no image pixels — upstream licenses prohibit hosting. The reconstruction toolkit (included in this repo) fetches images from their public/gated upstreams, verifies each file by sha256, and writes a complete local tree (11 synthetic conditions + reconstructed source images) identical in schema to the rest of the dataset.

Requirements 

pip install huggingface_hub datasets pyarrow pillow

Step 1 — Download this repository 

# Option A: snapshot download (recommended)
python -c "
from huggingface_hub import snapshot_download
snapshot_download('openconstruction/ConSynth-X', repo_type='dataset',
                  local_dir='./ConSynth-X')
"

# Option B: git clone (requires git-lfs)
git clone https://huggingface.co/datasets/openconstruction/ConSynth-X

Step 2 — Obtain upstream sources

Subset Upstream How to get
cs10k LouisChen15/ConstructionSite Public HF dataset — downloaded automatically by the rebuild script, no extra steps.
soda_voc SODA-D (Duan et al. 2022, Automation in Construction 142:104499) Request access from the SODA authors. Unpack so that <soda-d-root>/JPEGImages/hd08.jpg exists.
soda_ktsh SODA-ktsh (extended VL-annotated SODA variant) Request alongside SODA-D. Unpack so that <soda-ktsh-root>/images/Ktsh0005.jpg exists.

If you only have access to a subset of upstreams, pass --subsets to restrict reconstruction (e.g. --subsets cs10k).

Step 3 — Run the rebuild script 

python ConSynth-X/rebuild_consynthx.py \
    --release-root ConSynth-X \
    --out ./ConSynth-X-complete \
    --soda-d-root /path/to/SODA-D/VOC2007 \
    --soda-ktsh-root /path/to/soda-ktsh
    # --hf-token $HF_TOKEN  # only if cs10k is gated for your account (currently public)

The script will:

  1. Fetch each upstream image (cs10k auto-downloaded via HF datasets; SODA from local dirs you provide).
  2. Verify sha256 of every reconstructed image against original_manifest.parquet. Any mismatch or missing file is skipped and logged — an unverified image is never written.
  3. Merge the verified original rows with the 11 synthetic condition parquets (copied verbatim from the release root).
  4. Integrity-check row counts (cs10k 1,586 × 11 = 17,446 synthetic + 1,586 source; soda_voc 608 × 11 + 608; soda_ktsh 915 × 11 + 915).

Exit codes: 0 = fully clean; 1 = some mismatch/missing (partial output, see log); 2 = required upstream root missing; 3 = Mode B not implemented.

Because GT bounding boxes are stored as normalized [0, 1] coordinates, they apply without drift as long as the reconstructed image dimensions match the manifest — which the sha256 check guarantees.

Toolkit files (repo root)

File Purpose
rebuild_consynthx.py One-command reconstruction orchestrator
original_manifest.parquet Per-image provenance: upstream pointer, sha256, width/height, EXIF flag (3,109 rows)
README_reconstruct.md Detailed reconstruction guide (upstreams, exit codes, sha256 rationale)
UPSTREAM_ATTRIBUTION.md Upstream citations and license terms

Versions

Branch Release Rows Notes
main v2-strict (DINO ≥ 0.85) 34,199 11 synthetic conditions × 3,109 source scenes (current)

Related (relaxed-gate) releases of v2 family:

  • Ben11304/ConSynth-X-v2-strict — mirror of main (kept for citation links).
  • v2 (DINO ≥ 0.75, 99,828 rows) and v2-unfiltered (216,000 rows) — internal; reach out for access.

License and provenance

  • Synthetic conditions (11 splits): CC BY-NC-ND 4.0 — ConSynth-X's own output.
  • Source image GT (bounding boxes, captions in the original split): derived from upstream datasets; inherits their respective licenses:
    • cs10k GT → CC BY 4.0 (Construction-Site dataset, LouisChen15).
    • soda_voc / soda_ktsh GT → SODA research-only license. If you publish work using SODA-derived annotations, please cite:

      Duan, R., Deng, H., Tian, M., Deng, Y., & Lin, J. (2022). SODA: A large-scale open site object detection dataset for deep learning in construction. Automation in Construction, 142, 104499. https://doi.org/10.1016/j.autcon.2022.104499

  • Reconstructed original image pixels: inherit the upstream dataset license, not ConSynth-X's. See UPSTREAM_ATTRIBUTION.md for full citation and access terms per upstream.

Citation 

@dataset{consynth_x_v2_strict_2026,
  title  = {ConSynth-X v2-strict: DINO-gated paired synthetic augmentation for construction vision robustness},
  author = {Duong, Viet Huy and Xiong, Ruoxin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/openconstruction/ConSynth-X}
}

Acknowledgements

OpenConstruction Open Science Initiative · Construction Management Lab, College of Architecture and Environmental Design, Kent State University. Built atop Construction-Site, SODA, Depth Anything V2, InstructPix2Pix, CycleGAN-Turbo, and FLUX.1-Fill-dev.
Downloads last month
48
Total file size:
12.6 GB