ConSynth-X — Download & Usage

ConSynth-X release v1 is a multi-condition synthetic-augmentation benchmark of 122,156 construction-site images across three sub-datasets, packaged as Parquet shards with embedded JPEG bytes. See dataset_description.pdf for full content, generation pipelines, schema, and known issues.

Sub	Rows	Annotation	HuggingFace config
cs10k	44,105	3-class detection + 4 rule-violation categories + image attributes	cs10k
soda_voc	56,021	15-class detection	soda_voc
soda_ktsh	22,030	Caption-only (5 captions/image)	soda_ktsh

---

Condition grid preview

Each row shows the same source image transformed across the released conditions (original, rain_light, rain_heavy, snow_light, snow_heavy, fog_light, fog_medium, fog_heavy, night / day2night, small). Empty cells = condition not yet generated for that source id.

cs10k — construction-site detection (3-class + rule violations)

soda_voc — SODA 15-class detection

soda_ktsh — SODA caption sub (5 captions/image)

---

Quick start

1. Install

pip install datasets pyarrow pillow huggingface_hub

2. Stream a single config (no full download)

from datasets import load_dataset

ds = load_dataset("Ben11304/ConSynth-X", "cs10k", split="train", streaming=True)
for row in ds.take(3):
    print(row["image_id"], row["condition"], row["condition_labels"])

3. Download a single config locally

from datasets import load_dataset

ds = load_dataset("Ben11304/ConSynth-X", "soda_voc")
print(ds)            # DatasetDict with 'train' split (or whatever HF resolved)
print(ds["train"][0]["image_id"], ds["train"][0]["condition"])

4. Snapshot-download specific shards via huggingface_hub

For granular access without datasets:

from huggingface_hub import snapshot_download

local = snapshot_download(
    repo_id="Ben11304/ConSynth-X",
    repo_type="dataset",
    allow_patterns=["soda_voc/parquet/fog_*/*.parquet"],   # only fog
    local_dir="./consynthx",
)
print(local)

CLI form:

huggingface-cli download Ben11304/ConSynth-X \
    --repo-type dataset \
    --include "cs10k/parquet/rain_heavy/**" \
    --local-dir ./consynthx

5. Read directly with PyArrow (no datasets)

import pyarrow.parquet as pq
from io import BytesIO
from PIL import Image

t = pq.read_table("consynthx/soda_voc/parquet/fog_heavy/fog_heavy.parquet")
print(t.schema)
row = t.slice(0, 1).to_pylist()[0]
img = Image.open(BytesIO(row["image"]))
print(row["image_id"], img.size, row["condition_labels"])

---

Schema

Common fields across all three subs:

Field	Type	Description
image	bytes	JPEG bytes, byte-identical to the source — never re-encoded
image_id	string	Unique within sub
source_id	string | null	Original-image identifier (joins back to clean source)
source_dataset	string	"cs10k" | "soda_voc" | "soda_ktsh"
condition	string	Primary condition label (fog_heavy, rain_light, ...)
condition_labels	list[string]	All applicable labels (multi-label filter)
objects	list[struct]	{class_id: int, class_name: string, bbox: [x1,y1,x2,y2] normalised}
pipeline	struct	{method, checkpoint, prompt_template, params} (params is JSON string)
quality_scores	struct | null	{dino_sim, ssim, clip_sim, lpips}; null when not yet computed
quality_alert	bool | null	dino_sim < 0.75; null when DINO unavailable

Sub-specific extensions:

• cs10k adds image_attributes (caption, illumination, camera_distance, view, quality_of_info) and rule_violations (list of {rule_id, bbox, reason}).
• soda_ktsh adds captions (list of 5 natural-language captions per image; populated for 21,789 / 22,030 (98.9%) of rows. The remaining ~1.1% are rows whose image_id does not appear in any caption-bearing source arrow and are released with captions = []).

Bounding-box conventions:

objects[].bbox is [x1, y1, x2, y2] normalised to [0, 1] — resolution-agnostic, invariant under resize. Convert to pixel [x, y, w, h] per row by multiplying by the image's (W, H) (decoded from image bytes).

---

Recipes

Decode an image

from io import BytesIO
from PIL import Image

img = Image.open(BytesIO(row["image"]))
img.save("preview.jpg")

Filter by condition

ds = load_dataset("Ben11304/ConSynth-X", "soda_voc", split="train")
fog = ds.filter(lambda x: "fog" in x["condition_labels"])      # multi-intensity fog
heavy_rain = ds.filter(lambda x: x["condition"] == "rain_heavy")

Use only high-quality augmentations

quality_alert = false means DINOv3 similarity ≥ 0.75 (synthetic stays close to source).

high_quality = ds.filter(lambda x: x["quality_alert"] is False)

# Strict — also drop rows whose DINO is missing:
high_quality_only = ds.filter(
    lambda x: x.get("quality_alert") is False
    and x.get("quality_scores") is not None
)

Detection: convert to YOLO format

import os
from io import BytesIO
from PIL import Image

def export_yolo(row, out_dir):
    img = Image.open(BytesIO(row["image"]))
    w, h = img.size
    img_path = os.path.join(out_dir, "images", f"{row['image_id']}.jpg")
    lbl_path = os.path.join(out_dir, "labels", f"{row['image_id']}.txt")
    os.makedirs(os.path.dirname(img_path), exist_ok=True)
    os.makedirs(os.path.dirname(lbl_path), exist_ok=True)
    img.save(img_path, "JPEG", quality=95)
    with open(lbl_path, "w") as f:
        for o in row["objects"]:
            x1, y1, x2, y2 = o["bbox"]            # normalised xyxy
            cx, cy = (x1 + x2) / 2, (y1 + y2) / 2
            bw, bh = x2 - x1, y2 - y1
            f.write(f"{o['class_id']} {cx:.6f} {cy:.6f} {bw:.6f} {bh:.6f}\n")

cs10k: caption + scene attributes

attrs = row["image_attributes"]
print(attrs["caption"])
print(attrs["illumination"], attrs["camera_distance"], attrs["view"])

cs10k: rule violations

for rv in row["rule_violations"]:
    print(f"rule {rv['rule_id']}: {rv['reason']}")
    if rv["bbox"]:
        print(f"  bbox (normalised xyxy): {rv['bbox']}")

soda_ktsh: 5 captions per image

for caption in row["captions"]:
    print("-", caption)

Decode pipeline params

pipeline.params is a JSON-encoded string for forward-compatibility:

import json
params = json.loads(row["pipeline"]["params"])
print(row["pipeline"]["method"], row["pipeline"]["checkpoint"])
print(params)

---

Citation

@dataset{duong2026consynthx,
  title  = {ConSynth-X: A Large-Scale Synthetic Construction-Site Image Dataset
            for Challenging Field Conditions},
  author = {Duong, Viet Huy and Xiong, Ruoxin},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Ben11304/ConSynth-X}
}

Please also cite the upstream datasets:

@misc{chen2025vlm,
  title  = {Are Large Pre-trained Vision Language Models Effective
            Construction Safety Inspectors?},
  author = {Chen, Xuezheng and Zou, Zhengbo},
  year   = {2025}
}

@article{duan2022soda,
  title   = {SODA: A large-scale open site object detection dataset for
             deep learning in construction},
  author  = {Duan, Rui and Deng, Hui and Tian, Mao and Deng, Yichuan and Lin, Jiarui},
  journal = {Automation in Construction},
  volume  = {142},
  pages   = {104499},
  year    = {2022},
  doi     = {10.1016/j.autcon.2022.104499}
}

---

License

• Dataset (Parquet shards with embedded JPEG bytes): CC BY-NC 4.0
• Generation and repack code: Apache-2.0 (separate repository)

Source-data attributions: Construction Site 10k (LouisChen15) and the SODA dataset family (Duan et al., Automation in Construction 2022; SODA-KTSH extension, Deng et al. 2025). Downstream redistribution must respect both the ConSynth-X licence and the upstream source licences.

See dataset_description.pdf for the full data card, including known issues and ongoing DINO compute coverage.